About this book...

This book is a must read for anyone involved in creating a product;

management, testers and developers. Use it to change the way you think
projects should be run and close those communication gaps.
Toby Henderson
Have you ever wanted a better way to communicate, clarify and satisfy
business requirements? Wouldn't it be great if those requirements
evolved along with the software, always consistent and clear? And those
requirements helped drive development so that we knew when we were
done? With clarity, Gojko describes an elegance and effective way of
achieving this with the whole team: Inventing, thinking and
communicating with specific, insightful examples that also serve as
acceptance tests.
Rick Mugridge, www.rimuresearch.com, and lead author of Fit for
Developing Software, Prentice Hall, 2005.
Effective software development is all about good communication and
this book explains the toolkit that allows us to do this effectively. Worth
reading whatever your role is.
Mark Needham, http://www.markhneedham.com/blog/
Whether youre new to testing, new to agile, or an old pro at one or both,
youll experience aha moments that will inspire your team as you read.
This book will challenge some of your preconceived notions and make
you think. It paves the way for people in different roles, such as business
analysts, QA engineers and developers, to adapt to a more productive
agile approach. From practical ways to improve communication with
customers, to helpful examples of useful test tools, this book is a major
addition to our agile testing knowledge base.
Lisa Crispin, co-author, Agile Testing: A Practical Guide for Testers
and Agile Teams, Addison-Wesley Professional 2009
Gojko addresses an underrated point: that Test-Driven Requirements,
or Executable Requirements, are not about tools, automated tests, or
even professionalism. They are about communication. I wish each of
my colleagues and clients had a copy of his book, and maybe that fact
will be made just a little bit clearer to them.
Eric Lefevre-Ardant, Agile Coach and Developer,
http://ericlefevre.net/

I wish that the book had been available a few years ago when the
company I was at (and myself) were trying out agile. Could have been
a lot easier and more successful if wed read it.
Philip Kirkham
If youve tried agile acceptance testing youll know that as well as being
really exciting its also incredibly difficult. Luckily we now have a book
that helps guide us through the many tricky choices that we face,
practical and pragmatic advice that even the most experienced agile
developer should be aware of.
Colin Jack, Senior Software Developer, FNZ
As a tester, I welcome any opportunity to increase shared understanding
of requirements and expectations - our team will be relying on this book
to guide us as we begin our journey with agile acceptance testing.
Marisa Seal
You would be at least 6 moths ahead of the game in Agile QA by just
reading Gojkos book.
Gennady Kaganer, QA Manager at Standard and Poors
Gojko applies his experience to the practice of producing software that
is useful to end users. This is an important work in extending the
test-driven specification of software beyond individual units and into
the sum of the parts.
Bob Clancy, http://www.agiletester.net
Bridging the Communication Gap will not only bring you up-to-date
with the latest thinking about agile acceptance testing, but also guide
you as you put the ideas into practice. This book is packed with insights
from Adzics experience in the field.
David Peterson, creator of the Concordion acceptance-testing
framework
Im convinced that the practice of agile acceptance testing, done properly,
can make a dramatic improvement to both the communication with the
customer and the quality of the final product. This book is a solid
introduction to the subject, and represents the first attempt Ive seen to
survey the practice as a whole without focusing on a single tool or
technology.
Geoff Bache

Bridging the Communication

Gap
Specification by Example and Agile
Acceptance Testing
Gojko Adzic

Bridging the Communication Gap: Specification by Example

and Agile Acceptance Testing
Gojko Adzic
Copy-editor: Marjory Bisset
Cover design: Boris Marcetic
Published 5 Jan 2009
Copyright 2009 Neuri Limited
Many of the designations used by manufacturers and sellers to distinguish their products are claimed
as trademarks. Where these designations appear in this book, and the publisher was aware of a
trademark claim, the designations have been printed with initial capital letters or in all capitals.
The author has taken care in the preparation of this book, but makes no expressed or implied warranty
of any kind and assumes no responsibility for errors or omissions. No liability is assumed for incidental
or consequential damages in connection with or arising out of the use of the information or programs
contained herein.
All rights reserved. This publication is protected by copyright, and permission must be obtained
from the publisher prior to any prohibited reproduction, storage in a retrieval system, or transmission
in any form or by any means, electronic, mechanical, photocopying, recording, or likewise. For
information regarding permissions, write to:
Neuri Limited
25 Southampton Buildings
London WC2A 1AL
United Kingdom
You can also contact us by e-mail: [email protected]

Register your book online

Visit http://www.acceptancetesting.info and register your book online to get free PDF
updates and notifications about corrections or future editions of this book.

ISBN: 978-0-9556836-1-9

Table of Contents
About the PDF edition ............................................................... ix
Acknowledgements ................................................................... xi
About the author ..................................................................... xiii
Introduction ............................................................................. xv
Why should you care? ...................................................... xvi
Who is this book for? ...................................................... xvii
What will you get out of this book? ................................. xviii
What's inside? ................................................................. xix
Giving credit where credit is due ...................................... xxi
How this book is organised ............................................. xxii
I. The Communication Gap ........................................................ 1
1. The next bottleneck in software projects .......................... 3
The telephone game .................................................... 5
Imperative requirements are very easy to misunderstand .......................................................................... 8
Are obvious things really obvious? .............................. 9
A small misunderstanding can cost a lot of
money ...................................................................... 14
Fulfilling specifications does not guarantee
success ..................................................................... 15
Requirements are often already a solution ................. 18
Cognitive diversity is very important ......................... 20
Breaking The Spirit of Kansas ................................... 21
2. Finding ways to communicate better .............................. 25
Challenging requirements ......................................... 25
We really communicate with examples ...................... 26
Working together, we find better solutions ................ 28
Communicating intent ............................................. 29
Agile acceptance testing in a nutshell ......................... 31
So what does testing have to do with this? .................. 34
Better names ............................................................ 38
II. Building and Maintaining a Shared Understanding ............... 41
3. Specifying with examples ............................................... 43
How do you brush your teeth? ................................... 43
A practical example .................................................. 45

Bridging the Communication Gap

Realistic examples make us think harder .................... 48

Identifying important examples ................................ 50
Dealing with processing workflows ........................... 54
Working with business workflows ............................. 57
4. Specification workshops ................................................ 59
Running a workshop ................................................. 60
Workshop output ..................................................... 63
Sharing domain knowledge ....................................... 64
What specification workshops are not ....................... 68
How to keep the workshop focused ............................ 70
Dealing with open issues ........................................... 72
5. Choosing acceptance criteria ......................................... 75
Acceptance tests are the specification ........................ 75
Choosing the right set of examples ............................ 76
Distilling the specifications ....................................... 78
Specifying business workflows ................................... 80
Converting examples to tests ..................................... 82
Automating tests ...................................................... 86
Dealing with user interfaces ...................................... 91
Who should write acceptance tests? ........................... 95
What acceptance tests are not .................................... 97
6. Focusing development ................................................. 103
Building software to satisfy acceptance tests ............. 103
Suggesting new features .......................................... 105
Designing software based on acceptance tests .......... 106
Developing the glue between code and tests ............. 110
Use the project language consistently ....................... 111
Unit tests vs acceptance tests ................................... 113
Running acceptance tests ........................................ 115
Running manual tests ............................................. 117
7. Facilitating change ...................................................... 119
A live specification .................................................. 119
Keeping it live ......................................................... 120
Introducing changes ............................................... 123
Solving problems .................................................... 124
Keeping the software flexible ................................... 125
Keeping the specification flexible ............................ 126
III. Implementing agile acceptance testing ............................... 139
vi

8. Starting with agile acceptance testing ........................... 141

Agile acceptance testing in the context of the development process ........................................................... 141
Fitting into iterations .............................................. 144
Igniting the spark ................................................... 148
Don't mention testing ............................................. 150
Adopting in phases ................................................. 151
Choose a facilitator for the workshops ..................... 152
Hire a mentor ......................................................... 154
Avoid the cargo cult mentality ................................. 155
9. Planning with user stories ............................................ 159
User stories in a nutshell ......................................... 160
How are stories different from more traditional
methods? ................................................................ 161
Benefits of user stories ............................................ 162
Three Cs of user stories ........................................... 164
Project planning ..................................................... 165
User stories and acceptance tests ............................. 169
10. Tools of today ........................................................... 173
FIT ......................................................................... 173
Concordion ............................................................ 179
JBehave .................................................................. 181
TextTest ................................................................. 183
Selenium ................................................................ 187
11. Tools of tomorrow ..................................................... 193
Domain-specific languages ...................................... 193
Different user interfaces for different roles .............. 195
Propagating the effects of changes ........................... 196
Direct domain mapping .......................................... 198
Better editors .......................................................... 199
Better test organisation ........................................... 200
Visual workflow descriptions .................................. 201
IV. Effects on all of us ............................................................. 203
12. Effects on business analysts ........................................ 205
Benefits for business analysts .................................. 206
Challenges for business analysts .............................. 209
13. Effects on testers ........................................................ 217
The new role of the tester ........................................ 217
vii

Bridging the Communication Gap

Benefits for testers .................................................. 218

Challenges for testers .............................................. 222
14. Effects on developers ................................................. 229
New responsibilities for developers .......................... 229
Benefits for developers ............................................ 230
Challenges for developers ........................................ 232
A. Resources ........................................................................... 239
Online resources ............................................................. 241
Talks and videos ..................................................... 241
Presentations .......................................................... 242
Articles ................................................................... 242
Tools ...................................................................... 243
Mailing Lists ........................................................... 243
Index ...................................................................................... 245

viii

About the PDF edition

If you are reading this book in a print edition, you should know that
there is a PDF version available, and you might find it useful to get
an electronic copy as well. I've decided to make this book available as
a PDF so that the readers can get the most out of it. Judging from the
sales of my previous book, many people today prefer reading technical
texts in an electronic form so that they can search through the content,
copy and paste text and archive the material more easily.
If you are reading this as a PDF, please note that this material is
copyrighted and it is not for free distribution. I have no intention of
chasing people for redistributing the book, but if you have received
this PDF from someone and you found it useful, consider actually
purchasing your own copy. The PDF edition is relatively cheap and
your purchase will support similar efforts in the future.
Buying a PDF or a printed version entitles you to free PDF updates.
Register your copy and you'll receive a free PDF when the book is
updated in the future. For more information on the PDF version of
this book, and to register your copy, see:
http://www.acceptancetesting.info

ix

Acknowledgements
This book is a result of a small independent publishing effort, and as
such would not be possible without the help of many people.
I'd like to thank Antony Marcano, Bob Clancy, Colin Jack, David
Peterson, David Vydra, Eric Lefevre-Ardant, Gennady Kaganer, Geoff
Bache, Jennitta Andrea, Lisa Crispin, Marisa Seal, Mark Needham,
Melissa Tan, Mike Scott, Phil Kirkham and Rick Mugridge for all the
excellent suggestions, helping me keep this book focused and
providing insight into their views and experiences. Without you, this
book just would not be possible.
Marjory Bisset from Pearl Words again did a great job of copyediting
this book and ensuring that readers have a much more enjoyable
experience with it.
Finally, I'd like to thank Boris Marcetic from Popular for designing
the covers.

xi

xii

About the author

Gojko Adzic runs Neuri Ltd, a UK-based consultancy that helps
companies build better software by introducing agile practices and
tools and improving communication between software teams, stakeholders and clients. His programming story so far includes equity
and energy trading, mobile positioning, e-commerce, betting and
gaming and complex configuration management.
Gojko is the author of several popular printed and online guides on
acceptance testing, including Test Driven .NET Development with
Fitnesse and Getting Fit with .NET, and more than 200 articles about
programming, operating systems, the Internet and new technologies
published in various online and print magazines. He is the primary
contributor to the DbFIT database testing library which is used by
banks, insurance companies and bookmakers worldwide.
To get in touch, write to [email protected] or visit http://gojko.net.

xiii

xiv

Introduction
I am getting more and more convinced every day that communication
is, in fact, what makes or breaks software projects. Programming
tools, practices and methods are definitely important, but if the
communication fails then the rest is just painting the corpse. Complex
projects simply have no chance of success without effective communication.
This is a book about improving communication between customers,
business analysts, developers and testers on software projects, especially by using specification by example and agile acceptance testing.
Although these two practices are not yet popular, I consider them as
key emerging software development practices because they can
significantly improve the chances of success of a software project. (At
the same time, agile acceptance testing is one of the worst named
practices ever. For the time being, just forget that it has the word
testing in the name.) Agile acceptance testing and specification by
example essentially help us to close the communication gap between
different participants in a software project, ensure that they speak
the same language and build a truly shared and consistent understanding of the domain. This leads to better specifications the participants
flush out incorrect assumptions and discover functional gaps before
the development starts and build software that is genuinely fit for
purpose.
Ward Cunningham and Ken Auer used the basic ideas behind agile
acceptance testing to nail down what their users wanted in 1999.1
Almost a decade later, the practice is still used only by a small group
of early adopters. However, it has matured considerably. Judging
from recent conferences, magazine articles and blog posts, it seems
to me that interest is growing and the time is right for a wider group
to learn about and adopt it. You probably agree, which is why you
picked up this book.
1

See Don't just break software, make software[1]

xv

Introduction

Why should you care?

Agile acceptance testing and specification by example help business
people, software developers and testers communicate better and
understand each other. Each of these groups has different needs and
problems in a software project, so improved communication has
different benefits for them. In this section, I briefly list the the benefits
agile acceptance testing and specification by example will deliver for
each group. The rest of this book explains how to achieve these
benefits.
These are the most important benefits for product owners, business
analysts and project managers:
Developers will actually read the specifications that you write.
You will be sure that developers and testers understand the
specifications correctly.
You will be sure that they do not skip parts of the specifications.
You can track development progress easily.
You can easily identify conflicts in business rules and requirements
caused by later change requests.
You'll save time on acceptance and smoke testing.
From a developer's perspective, these are the most important
advantages of agile acceptance testing and specification by example:
Most functional gaps and inconsistencies in the requirements and
specifications will be flushed out before the development starts.
You will be sure that business analysts actually understand special
cases that you want to discuss with them.
You will have automated tests as targets to help you focus the
development.
It will be easier to share, hand over and take over code.
From a tester's perspective, these are the most important benefits of
agile acceptance testing and specification by example:
xvi

Who is this book for?

You can influence the development process and stop developers

from making the same mistakes over and over.
You will have a much better understanding of the domain.
You'll delegate a lot of dull work to developers, who will collaborate
with you on automating the verifications.
You can build in quality from the start by raising concerns about
possible problems before the development starts.
You'll be able to verify business rules with a touch of a button.
You will have a lot more time for exploratory testing.
You will be able to build better relationships with developers and
business people and get their respect.

Who is this book for?

This book is primarily intended for product owners, business analysts,
software developers and testers who want to learn about how to
implement agile acceptance testing and how to improve communication on their projects. It should also prove to be interesting to project
managers working on agile software projects or in the process of
migrating to an agile methodology. I wrote this book both for people
working in implementation teams and for people working on the
customer side, and for teams that practice agile development as well
as those who are migrating to an agile process. I use the term implementation team in this book as a name for the entire group of people
working on a project, from business analysts to developers, testers
and anyone else involved.
Although agile teams generally strive for less role-naming than those
practising traditional processes, I use the role names product owner,
business analyst, software developer and tester in this book and discuss
how each of these roles is affected by agile acceptance testing and how
they participate in writing specifications by example. Companies in
transition with teams coming from a more traditional background
will have all these roles clearly defined, sometimes isolated into
different departments, buildings or even countries. Most agile
programming literature is written by software developers for software
xvii

Introduction

developers, leaving testers and business analysts to fend for themselves

and try to work out what they should be doing. Agile acceptance
testing requires active participation from all project implementation
team members, including business analysts and testers, not just
developers. This is why I want to use these roles and names to
emphasise their involvement in the process. Contrary to common
belief, it is not unusual for larger agile projects to have dedicated
business analysts and testers. ThoughtWorks is generally accepted as
one of the leading agile software development companies and their
process has these roles clearly defined and uses these names.2
The roles of business analysts and testers on agile projects differ
somewhat from their roles on projects that use more traditional formal
methods, but this does not mean that the people behind these roles
do not exist. Roles are blurred a bit, but specialists still have to focus
on particular areas. If you are a business analyst or a tester, this book
should help you define more accurately what your job is and how to
make a more useful contribution to building software the agile way.

What will you get out of this book?

Primarily, this book will help you to get started and implement agile
acceptance testing and specification by example effectively. It explains
the principles and ideas behind the practice and helps you put it into
the wider context of your development process.
As with any other new and relatively radical idea, learning how to
apply it effectively in your own role is only part of the challenge. Agile
acceptance testing requires active participation and affects the jobs
of programmers, business analysts, testers and to some extent project
managers. Because of this, it also raises a lot of concerns and it is
unfortunately subject to many misconceptions. These problems are
typically fuelled by partial information, misunderstandings and fears
of change. One of my primary goals with this book is to dispel these

http://download.microsoft.com/documents/uk/msdn/architecture/architectinsight/2007/Lifecycle/LIF02-The-Agile-Template-for-VSTS.ppt.

xviii

What's inside?

misconceptions and address fears and issues that people often have
about agile acceptance testing.
It is also my intention with this book to challenge some established
ways of thinking in the industry. Agile acceptance testing breaks down
traditional boundaries around testing, requirements and specification
processes in a way that significantly improves communication on a
project. Specification by example is an approach to the writing of
specifications and requirements radically different from the established industry process. I will explain this in a lot more detail
throughout the book, but please note that if you come from a more
traditional software process background, you may need to put aside
the stereotypes that you are familiar with in order to grasp the ideas
and gain the full benefits.
This book will help you to discover how agile acceptance testing affects
your work whether you are a programmer, business analyst or a tester.
It will help you gain an understanding of this technique and offer
ideas on how to convince other team members and stakeholders to
use it. I hope that this book takes into account the various perspectives
of these different roles. It is intentionally not too technical or suited
only for programmers, because agile acceptance testing is not a
programming technique: it is a communication technique that brings
people involved in a software project closer. I consider myself
primarily a programmer and my personal biases and experiences will
surely show. I have, however, found a much broader perspective on
agile acceptance testing as a result of working with different teams,
consulting, mentoring and giving public talks. Discussing fears such
as losing jobs with testers over coffee, participating in requirements
gathering, working as a product owner and even as a business domain
expert on several projects has hopefully given me the understanding
necessary to represent the needs and views of the other roles.

What's inside?
I am very proud of the fact that I have helped several companies adopt
agile acceptance testing and specification by example and deliver

xix

Introduction

successful software over the last few years. This book is a summary
of my experiences, learnings and ideas from this journey. It is based
on projects that are now happily running in production and a series
of public and private seminars, talks and workshops that I organised
during this period.
This book contains a description of the process that I use today to
bridge the communication gap between business people and software
implementation teams. More importantly, it describes ideas, principles
and practices behind the process and their effects on people participating in software projects.
This book is not a detailed user manual for any acceptance testing
tool. Although I will describe briefly some of the most popular tools
in Chapter 10, I intentionally want to keep this book relatively nontechnical and absolutely not tool or technology specific. Often I hear
from managers that their teams looked into this or that tool and that
they did not like it a bit, so they rejected the whole idea of agile
acceptance testing. One of the main messages I want to convey with
this book is that the biggest benefit of agile acceptance testing is
improved communication and mutual understanding, not test automation. There is too much focus on tools today and in my opinion
this is quite wrong.
While you are reading this book, it's important to focus on the ideas
and principles. The process and the tools described in the book are
there just to assist. If you do not like a particular tool, find a better
one or automate things yourself in a way that makes sense for your
team. If you do not like some part of the process, adjust it or replace
it with something that makes more sense for your team. Once you
understand the underlying ideas, principles and practices, applying
them in your environment should be easy. Use the process described
in this book just as a guide, not as a prescription.
This book also does not have all the answers about agile acceptance
testing and specification by example. These two practices are currently
gaining a lot of momentum and generating a lot of new ideas. Because
there is very little published material on them at the moment, you

xx

Giving credit where credit is due

will see many more references to conference talks, blog posts and
online video clips than to books or articles. Apart from the practices
as I use them today, this book also describes some promising ideas
that I want to try out in the near future. It even contains some interesting views of other people that I do not agree with completely. Some
of them might prove to be interesting in the future with better tools
or be applicable in a different environment. In Chapter 11 I speculate
how the tools for agile acceptance testing tools might evolve. Blogs,
mailing lists and sites listed in Appendix A will help you continue the
journey and keep up-to-date with new advances.

Giving credit where credit is due

I make no claim and take no credit for inventing any of the ideas or
techniques that are described in this book. The techniques I describe
are based on ideas I have picked up on the way from books, conferences, blogs, discussion groups and colleagues. This book is an attempt
to collect different techniques that have emerged in the community,
show how they relate to each other and organise them in a one
coherent practice.
Agile acceptance testing is an evolution of the idea of customer tests
proposed by Kent Beck in which automated tests are used to verify
what the customers expect out of a software system. It also draws on
ideas developed by many others. Gerald Weinberg and Donald Gause
used the black-box requirements testing process to redefine the role
of tests in a project and relate them directly to requirements. Brian
Marick has done great work on bringing tests closer to the business
by clearly defining business-facing tests. He is also one of the key
promoters of the idea of specification by example, where realistic
examples are used instead of abstract requirements. Dan North's ideas
of behaviour-driven development bring acceptance testing closer to
business people. Ron Jeffries described the card-conversationconfirmation concept nailing down the basic ideas behind specification
workshops and linking acceptance tests to user stories. Joshua
Kerievsky's storytest-driven development expands on this to fit
acceptance testing into agile planning techniques. Jim Shore defined

xxi

Introduction

the describe-demonstrate-develop process as a way to use acceptance

tests iteratively to focus development. Lisa Crispin and Antony
Marcano were instrumental to promoting and improving agile testing
practices, tearing down the walls between testers and developers and
business people. Eric Evans came up with the idea of ubiquitous
language, used to avoid confusion caused by different jargons on a
project.
These ideas are supplemented by the practices and knowledge
embodied in Ward Cunningham's tool FIT and Object Mentor's tool
FitNesse, along with many other similar tools and extension libraries,
especially Rick Mugridge's FitLibrary and the learning that came out
of using these tools in commercial projects. David Peterson's focus
on specifications over scripting with Concordion is instrumental to
getting the most out of agile acceptance testing. A lot of the underlying
principles come from test-driven development, which in turn borrows
a lot of ideas from the zero quality control idea of Shigeo Shingo and
the Toyota Production System of Taichi Ono along with their counterparts in the software world, the lean software development ideas of
Mary and Tom Poppendieck. Effective application of agile acceptance
testing depends on various practices such as short iterations from
Scrum, Extreme Programming and other agile processes and the work
on user stories by Mike Cohn and others. Some ideas are even
borrowed from Prussian military tactics from the nineteenth century
and their current implementation in the US Army.
Agile acceptance testing is more than the sum of all these ideas and
practices. It is an approach to software development characterised by
a focus on communication and the creation of an understanding of
the domain shared by all implementation team members and
customers.

How this book is organised

In Part I, I present causes of the communication problems in software
projects, explain why the traditional model of building requirements
and specifications does not work and how other agile programming

xxii

How this book is organised

practices do not really solve the problem but only provide workarounds. Then I introduce agile acceptance testing as the solution to
these problems.
In Part II, I introduce the techniques and principles of agile acceptance
testing and explain how they work together to help us facilitate
communication and build better software.
In Part III, I talk about implementing agile acceptance testing in
organisations. I explain how this practice fits into the wider software
development process and how to start using it in your organisation.
I also briefly describe current popular tools for agile acceptance testing
and discuss what can we expect from future tools. This part also
includes a chapter on user stories, another agile technique that makes
the implementation of agile acceptance testing much easier.
In Part IV, I deal with the human side of this practice, explaining how
it affects our jobs and the way we work. I analyse the effects on business analysts, testers and developers. The chapter on the effects on
business analysts is also applicable to customers or other business
people involved in software projects. In this part we also revisit the
benefits listed in section Why should you care? on page xvi and see
how the principles and practices described this book deliver them.

xxiii

xxiv

Part I. The Communication Gap

Effective communication is the key to successful software projects. In this
part, we set the stage by looking at common communication problems
on software projects, analysing what happens when communication is
impeded and how we can improve the flow of information.

Chapter 1

The next bottleneck in

software projects
The world of software development went through quite a shake-up
in the last decade. Extreme programming and other early agile
development practices removed several huge bottlenecks in the way
we were used to working, allowing software teams to deliver faster,
better and react to changes more easily. Once these problems were
out of the way, people started noticing other issues in projects such
as planning and organisation. Looking for a better way to manage
projects, agile development teams adopted the Scrum methodology.
A whole range of related techniques emerged to remove planning and
organisational bottlenecks. Agile development and management
practices are now mature and widely accepted and I feel confident
that the technical part of building software is no longer a problem.
The most important issue now is making sure that we know what we
need to build, but this problem is a whole lot harder than it sounds.
Of course, this problem is nothing new. Fred Brooks wrote long ago
that The hardest single part of building a software system is deciding
precisely what to build[2]. Practices used in the creation of traditional
system requirements documents and specifications have changed
very little over the last decade. Some techniques such as user stories
touch upon what we should do to get better specifications and
requirements, but they fall short of completing the task to the end.
The process of eliciting requirements and communicating them
throughout the process to everyone involved is often just described
in vague terms and left to teams to work out for themselves. With
programming practices and project planning now pretty well sorted,
this issue is becoming more and more apparent and many teams
realise that they struggle with it. This problem is the next bottleneck
to be removed from the software development process.

The next bottleneck in software projects

A nice thing about software development the agile way is that we can
easily go back and adjust the system, but making changes is not as
cheap as most programmers would like it to be. It costs a lot of money
and time. For those of you who would now suggest that using agile
practices makes this cheap, don't forget that you are only talking about
the technical cost. Agile practices help us cope with clients changing
their minds for whatever reason, so that in a sense we are hitting a
moving target. There is a huge difference between this and missing
a still target and then going back to have another go at hitting it. Even
if the requirements don't change, there is still a risk that the project
can miss a target if we don't solve the communication problems.
Although agile practices help a lot with reducing the risk of failure,
they should not and cannot be used to cover up for a project that
simply does not deliver what it was supposed to. Disappointing a
client is never good, agile or not agile. Gerald Weinberg and Donald
Gause suggest that the difference between disappointment and delight
is not a matter of delivering software, but how well the delivery
matches what clients expected.1 This was true twenty years ago when
they wrote Exploring Requirements[3], and it still holds true today.
Matching what clients expect is still a problem, mostly because of
communication issues. Individual effects of small communication
problems are very hard to detect, so they do not become apparent
instantly. Such problems are reflected in lots of small things not
working as expected or implied features that simply do not get
delivered. Individual issues may have small effects on the project, but
their cumulative effect is huge. The reason why changes in development practices in the last ten years have not solved this problem is
that most of these changes were driven by developers and I do not
believe that this particular issue is a development problem at all. It is
a communication problem involving all participants in the implementation team. This is why there is no development practice that can
solve the problem, whether or not it demands the involvement of
customer proxies and business people.

See chapter 18 of Exploring Requirements[3]

The telephone game

We all need to agree on what the target is, even if it moves, and make
sure that we all have the same understanding. And by we I mean all
participants in the process from stakeholders to domain experts,
business analysts, testers and developers. The path to success is to
ensure that these small communication problems get rooted out
instead of accumulating, so that the message gets delivered correctly
and completely.

The telephone game

The traditional model of gathering requirements and building
specifications is based on a lot of formalising, handing over and
translating. Business analysts first extract knowledge about requirements from customers, formalising it into specifications and handing
it over to developers and testers. Developers extract knowledge from
this and translate it into executable code, which is handed over to
testers. Testers then take the specifications, extract knowledge from
them and translate it into verification scripts, which are then applied
to the code that was handed over to them by the developers.
In theory, this works just fine and everyone is happy. In practice, this
process is essentially flawed and the usual result is a huge difference
between what was originally requested and what gets actually
delivered. There can be huge communication gaps at every step.
Important ideas fall through these gaps and mysteriously disappear.
After every translation, information gets distorted and misunderstood,
magnifying the degree of wrongness of the delivered system. A tester's
independent interpretation might help to correct the false interpretations of developers, or it might very well be a completely different
misinterpretation of the system requirements.
With agile development processes, the feedback loop is much shorter
then in a traditional process, so problems get discovered quickly.
However, if agile acceptance testing is not applied, even with other
agile practices in place, there is still a lot of scope for mistakes. Instead
of discovering problems, we need to work out how to stop them from
appearing in the first place.

The next bottleneck in software projects

People seem somehow surprised by this effect even though most of

us have encountered it and used to amusing effects in childhood.
Antony Marcano draws a parallel between the traditional software
development process and the telephone game.2 In the telephone game
(called Chinese whispers in the UK) a group of children stand in line
and then the first child whispers a phrase or sentence to the next child.
The second child whispers what it heard to the third one and so on.
The last child in the line says aloud the phrase or sentence that it
heard. It is often significantly different from the original phrase.
Although these cumulative differences may have been amusing when
we were children; they are not so funny when it comes to solving real
problems that obstruct people in doing their jobs. What happens with
the traditional approach is that quite a few things get lost in translation. Even when people do read big documents carefully, they don't
remember all the details. Most people just digest the text looking for
key ideas and guidelines, disregarding everything else.
Relying on people to remember all the details is futile. This is why
features rarely get implemented completely and correctly at the first
attempt, so that the project relies on testers running through the
specification documents and comparing them with what has been
developed. This again poses a challenge since testers often get involved
only when the party is over and they are expected to understand
instantly a complex software system without taking too much of the
precious developers' time on explaining things. As tight deadlines
are a rule rather than an exception these days, testers don't have a lot
of time to learn about the development and system specifications.
They are often expected to approve a release in just a few days. This
makes it hard for them to really understand what they are testing and
severely undermines the effectiveness of testing.
Although big specification documents give the impression of constituting complete and comprehensive descriptions of requirements, in
practice (at least the ones I have seen) they often leave out some details
that need to be worked out later via e-mail or in a whole series of other
documents. The original specification is, in such cases, out of date as
2

http://www.testingreflections.com/node/view/7232

The telephone game

soon as the development starts. Even with perfectly correct, complete

and detailed specifications, there is still a possibility that requirements
will change during development. This is a rule rather than an exception for longer projects. Changes get implemented directly without
updating specifications, leading to differences between code and
system documentation. In this case the specifications are technically
incorrect. So if the testers identify a difference between the system
and the original documents, this does not necessarily mean that they
found a problem. This makes the job of testers even harder, because
they cannot rely completely on the specifications.
Such differences between what was asked for and what was developed
may also come from unclear or inconsistent requirements. Code is
unforgiving with regards to specifying how something should work.
On this level, we have to define exactly how the system should behave.
Once developers start writing code, gaps in functionality become
obvious. Unfortunately, getting to domain experts or business people
who wrote the specifications at this point may take a while. In the
best case, developers need to get the person on the phone or chase
her by e-mail. If the person in charge is an external customer or an
executive project sponsor, they might not be readily available to talk
and it might take a few days to get the right information. In some
extreme cases, it is nearly impossible. During a workshop I organised
for a large media company, one of the developers said that their
business analysts actually refuse to discuss requirements after they
were handed over to development. According to their process, the
job of business analysts was done at that point.
Because developers need to specify how the system works in the newly
identified case and they need to seek out this information, the time
to complete a piece of code increases from several minutes to several
hours or several days. When the gaps are finally cleared up, the
original specification document again may not reflect the full reality.
There is at least some new functionality not covered by the document,
or it may even conflict with the original requirements.
Later on, when requests for changes and improvements start coming
in, especially if the product has several clients with different require-

The next bottleneck in software projects

ments, changes may break earlier rules or introduce inconsistencies

into the system. Depending on regression test coverage, this problem
might surface or be noticed only in production.
The problem is, of course, that not even the brightest among us can
keep track of all the previous requirements and all the changes in
their minds for a long period of time. Realistically, it's not sane to
expect someone to remember all 500 pages of a specification document
and instantly spot that a new change request might break an earlier
rule. This becomes even harder when the people involved in the
original project move on to other things and are replaced. With big
requirements documents, consistency can be a problem even in the
initial phase, since they might be the result of several days of discussions with different people, and are sometimes written by several
people.

Imperative requirements are very

easy to misunderstand
With agile development processes, it is less likely that business analysts
will produce great long documents that caused a small forest to be
cut down. Requirements are typically broken down and iteratively
constructed and communicated. This does not really make a huge
difference with regards to the communication problems that I want
to attack. Without agile acceptance testing, agile processes will
potentially help us catch a problem sooner because on-site customers
provide feedback, but they do not prevent the problem from
happening. Requirements are simply subject to interpretation and
quite often to misinterpretation. Imperative requirements, which
command what people should do without demonstrating it, are
especially bad as they are seemingly precise but leave quite a lot of
potential for mistakes. We each have our own assumptions that affect
the way we understand these statements. The fact that we have heard,
read and understood something in English does not necessarily mean
that we understood it in the same way.

Are obvious things really obvious?

The story about a series of experiments conducted by Lawrence G.

Shattuck with four active battalions of the US Army,3 quoted in Gary
Klein's book Sources of Power[4], illustrates this problem nicely. The
researchers attended a military exercise and listened in while the
commanders were giving orders to their teams. During the exercise,
they wrote down everything that the teams did. After the exercise,
they discussed the actual actions of the teams with the commanders.
The results of the research were that actions on the ground matched
commanders' expectations completely only in 34% of the cases. While
I was reading it, it struck me how this story completely relates to my
early experiences in software development. The commanders are
customers or business analysts, the orders are requirements or
specifications and the teams on the ground are developers and testers.
We all sit in the same room, listen to the customers, agree on the
requirements and specifications and then go away and develop
something that often does not match completely what the customers
want. I don't know of any similar experiments actually being
conducted with software development teams, but I would not imagine
that the result is a lot different. In a sense, the story about the same
thing happening in the US Army is comforting because it means that
we are not the only ones plagued by misunderstandings and wrong
assumptions. Unfortunately, this comforting feeling quickly goes
away if you think about the fact that these guys have the biggest guns
in the world and that they misunderstand orders twice in every three
times.

Are obvious things really obvious?

A very serious problem with requirements is taking obvious things
for granted. In Exploring Requirements[3], Gause and Weinberg
described an experiment from one of their workshops that shows how
even the simplest things can be misinterpreted. They showed a sevenpointed star picture to attendees before a seminar, then held a lecture
and let people have a coffee break. After the break, they asked the
attendees to tell them how many points the picture had. They received
3

http://www.au.af.mil/au/awc/awcgate/milreview/shattuck.pdf

The next bottleneck in software projects

quite a wide spread of results. According to Gause and Weinberg,

everyone knows what a five-pointed star looks like but their star was
more unusual, so people remembered it differently and recalled
different images, especially after the coffee break. This accounted for
the spread. People also understood the task differently, which
explained clusters in the answers.
In August 2008, I repeated this experiment,4 focusing only on
differences in interpretation. During a talk, I handed out index cards,
put a picture of a standard five-pointed star on the projector and
asked the attendees to write down how many points they saw in the
picture. I used an image of a familiar figure (Figure 1.1) and it was
on display while people wrote down their answers, so that observation
and memory problems would not affect the answers. Any differences
in answers could only be caused by people interpreting the task
differently. After some initial reluctance and explaining that the
question might sound stupid and obvious, but people should answer
it anyway, I got more than 40 cards back.
An interesting thing about this experiment is that most people are
absolutely certain that there is an obvious answer, because the question
seems so simple. The real problem is that although people have an
obvious single logical answer in their heads (the only possible
answer), this answer differs from person to person (Figure 1.2).
Twenty five people voted for ten points in the star, counting inner
and outer points. The second most popular answer with seven votes
was five points, where people counted just the outer points. Six people
voted for fourteen one of the attendees explained that this is probably the ten points on the star and the four corners of the picture.
Note that there was no border on the picture, but some people decided
to count the edges of the screen as well. There was a single vote for
nine points, which can theoretically be explained by five outer points
and four corners of the picture. A reader of my blog later suggested
that eleven would be a natural answer for someone with a vector
graphics background, who would consider that the end point and
starting point are different points even though they are physically in
4

http://gojko.net/2008/08/29/how-many-points-are-there-in-a-five-point-star/

10

Are obvious things really obvious?

the same place. The two votes for fifteen are still a mystery to me.
Bob Clancy suggested that this might be a classic example of the
original problem being restated by the individual and than the restated
problem solved: if the star is drawn with lines crossing in the inside
of the star and then a point counted wherever two lines intersect,
people might count the points in the inner pentagon twice (although
they are the same as the five inner points).

Figure 1.1. A standard five-pointed star... but how many

points does it have?

Some readers are now probably asking themselves what is the right
answer. In general, there is no right answer. Or more precisely, all of
these answers are correct, depending on what you consider a point.
In software projects, on the other hand, there is a single correct answer
in similar situations: the one that the business people thought of. And
for the project to turn out just the way that customers want, this
answer has to come up in the heads of developers and testers as well.
This is quite a lot of mental alignment. Forty cards are not a sample
large enough for statistical relevance, but this experiment has
confirmed that even a simple thing such as a familiar image and a
straightforward question can be interpreted in many different ways.
11

The next bottleneck in software projects

Figure 1.2. Results of my poll on points in the star

Imagine that you are part of a software project that calculates prices
for gold-plating various metal pieces. This may be a contrived
example, but let's keep it simple for now. One of the requirements,
in its classical imperative form, states that the system shall let the
users enter a diameter and the number of points of a star-shaped
metal piece and calculate the price of gold-plating based on materials
(total piece surface area using prices in appendix A) and complexity
(number of edges using prices in appendix B). We have the prices
and formulas precisely defined in imaginary appendices and the
business analyst has spent a lot of time getting these absolutely clear
with the customers, because this is where the money is. Developers
should be able to work out the rest easily from the number of points
in the star and the diameter and testers should be able to compare
the test results to expected results easily. After all, everything is
specified precisely and we only have to do a bit of elementary
geometry. Right? Well, not exactly.
When requirements finally come to development, things become
much more precise because someone actually has to write the code.
Bertrand Russell wrote in The Philosophy of Logical Atomism[5] that
Everything is vague to a degree you do not realise till you have tried
to make it precise. On most projects even today, writing code is the
first time that we try to make the solution really precise. At this point,
a developer may have the same understanding of a point as the busi12

Are obvious things really obvious?

ness person making the request, but I would not bet on it. Work out
the probabilities from the experiment, and you'll get about a 39%
chance for this to happen. A tester will need to verify the result which
asks for another mind alignment and brings down the probabilities
to 20%. Again, I don't claim that this number is statistically significant
and describes a general success ratio, but whatever the precise figures
the probability falls exponentially with the number of participants
that need to have their minds aligned. Problems like this often don't
come to light before development because they are subtle and hidden
behind things perceived as more important such as the price of goldplating per square inch. We think that we don't have to be precise
about things that everyone understands during analysis, because it's
common sense how to draw a star with a number of points. The rules
for the surface area should be clear from basic geometry. Let's
disregard all the weird answers and consider just the fact that some
people in the experiment counted only outer points, and some counted
both outer and inner points. How would you test whether the system
works correctly if 12 was given as the number of points? Is the correct
star the one on the left or one on the right in Figure 1.3?

Figure 1.3. Which one of these has 12 points?

Things like this, where we feel familiar with the concept and implicitly
think that others have the same understanding of it as we do, are one
of the core causes of missing the target in software projects.

13

The next bottleneck in software projects

A small misunderstanding can cost

a lot of money
The previous example was invented to demonstrate the point and
arguably the misunderstanding is not a big one the prices will be
worked out correctly once everyone agrees on what a star with 12
points looks like. But the devil is in the detail. Wrong assumptions
and misunderstandings that we have about the domain might not be
huge, but they can still have a huge impact on the software that we
deliver. Here is a real world example of how a tiny misunderstanding
can cause a lot of pain.
Online poker is big business in UK at the moment, so loads of
companies want to grab a piece of the action. Because poker is
essentially a group game, potential newcomers are presented with a
chicken-and-egg problem. In order for your poker system to be
interesting to new players, it has to have lots of other players online.
It is very hard to start up a network of players initially. Instead of
attempting this, lots of UK bookmakers just make a deal with an
existing poker network and then re-brand the software, instantly
giving their customers many potential opponents and sharing a piece
of the profit of the whole network. UK punters have accounts in
British pounds but most poker networks work with US dollars as
chips. So the system required a screen that allowed players to convert
between UK pounds and US dollars. The requirement for this was to
round the converted value to two decimal places, as a poker chip was
worth one cent and fractions of cents were not supported. People who
have worked on financial systems in the past probably already guess
what happened.
This worked fine until one day someone found out how to make
money out of it. For the sake of the story, let's say that the exchange
rate was 0.54 pounds to a dollar. So 0.01 pounds would convert to
0.02 dollars, but a single cent would convert to a single penny 0.0054
rounded to two decimals is 0.01. Because of the rounding, you could
convert a single penny into two cents, then convert the first cent back

14

Fulfilling specifications does not guarantee success

to a penny, then convert the second cent into another penny and end
up with twice the money you started with. Yes, it is just one penny
more, but the guy who worked this out wrote a script to do the dirty
job and apparently took more than ten thousand pounds before the
fraud was discovered.
When the news about this broke, business people argued that the
amount should have been rounded down and that the developers
should have known this, but the developers argued that they received
a request to round to two decimals without any specifics. The
requirement to round to two decimals sounds obvious and unambiguous, just as did the question about the number of points in the star.
In any case, the blame game does not solve the problem. We need to
prevent problems like this by ensuring that developers and business
analysts have the same understanding of rounding to two decimals.
The question needs to be raised before the development starts, not
after it is in production, when someone finds a hole in the system.

Fulfilling specifications does not

guarantee success
The communication gap on software projects doesn't only cause
people to understand things differently. It also causes them to focus
on unimportant issues. This is a huge source of problems. I have seen
several projects where people wasted enormous amounts of effort
building features that they thought were important, but in fact missed
the real goals. Working from specifications does not guarantee that
the project will deliver the desired business value.
Gilles Mantel organised a workshop entitled Test-Driven Requirements: beyond tools5 at the Agile 2008 conference held in Toronto in
August 2008. At the workshop, four volunteers acted as a customer,
a product owner, a developer and a tester for a simple construction
project involving children's play bricks and dominoes. (For the sake
5

http://testdriveninformation.blogspot.com/2008/08/material-of-tdr-workshop-at-agile2008.html

15

The next bottleneck in software projects

of the story, if you don't know what a product owner is, think of it as
a business analyst.) There were two sets of construction bricks and
dominoes, placed on two halves of a large table with a large white
paper screen blocking the view in between, as shown in Figure 1.4.6
A person sitting at one side of the table could not see anything on the
other half. The aim of the workshop was to demonstrate how the
requirements and testing processes fail when communication is
impeded, even with something as simple as a domino construction.
At the start of the workshop, the product owner, the developer and
the tester left the room. Mantel put together a simple construction
on one half of the table with the customer, aligning the dominoes so
that they all fell when a small lead ball was rolled down one of the
bricks and bounced against another brick then the first domino. There
were lots of other bricks in the construction, but the primary business
goal was to make all the dominoes fall when the ball was rolled.
The product owner then walked into the room, looked at the
construction and discussed it with the customer. Then the developer
and the tester walked into the room. The developer was placed at the
other half of the construction table, unable to see the original
construction. The tester was placed on a different table, facing the
construction table backwards and unable to see what was going on.
This was intended to simulate the situation where a product owner
acts as a customer representative and testers have no influence over
development.
The product owner then started explaining to the developer how to
replicate the construction. The developer started building with the
bricks on his part of the table and the tester listened in to the
conversation. The developer was allowed to ask any questions and
the product owner was allowed to explain the construction in any
way he saw fit, but the tester was not allowed to ask any questions
during the construction. She just listened in and took notes. The
product owner led the developer by explaining the shape and relative
positions of different building blocks. They discussed at length
6

The photograph is Eric Lefevre, used with his permission

16

Fulfilling specifications does not guarantee success

differences in colour and number of blocks. When the developer had

finished building the system, the tester was asked to validate what the
developer had built and approve or reject the release, without looking
at the original construction. Based on her notes, the tester found a
few issues and refused to approve the release. Gilles Mantel then
performed the only test that was really important he rolled the ball
down the first brick. It did not bounce off the second brick and the
dominoes did not fall.

Figure 1.4. Domino construction exercise at Agile 2008

The interesting thing was that the product owner never asked the
customer about the goal of the construction, so he was not able to
communicate anything in this respect. This was not forbidden by the
rules of the game, it just never happened. When the exercise ended,
the customer said that the ordering of dominoes, alignment of bricks
by colour and all the extra bricks on the table were not really
important and that he would have accepted the construction if these
were different providing the dominoes fell after the ball was rolled.

17

The next bottleneck in software projects

Unfortunately, the product owner and the developer wasted most of

the time discussing exactly these issues and the tester validated the
construction based on this discussion.
This exercise demonstrated very clearly what happens with the
traditional requirements process, when the important business goals
are simply lost in translation and lots of effort is wasted on irrelevant
details. The developer had no real chance of fulfilling the business
goal because he had no idea what it was.
The tester was just asked to approve or reject the release without really
understanding what the project was about, based just on the
specification put together by the product owner. So her contribution
was left to pure chance as well. She did reject the release, but for the
wrong reasons. In this case, it was lucky that the developer did not
build the construction exactly as the product owner required, otherwise the tester would have approved the release. The result would
have been correct in terms of the specifications, but would have
completely missed the business goal.
Describing how and what but not why left the success of the project
to pure chance. Instead of producing traditional specifications and
requirements, we should really focus more on getting the communication right. Involving developers and testers from the start,
communicating business goals to everyone and removing communication obstacles is the way we can take control of projects, and not
leave success or failure to pure chance.

Requirements are often already a

solution
Classical project requirements focus on what but not on why. They
are effectively a proposed solution to a problem, but do not tell us
what the problem actually is. Customers and business people who
specify requirements generally have a very superficial view of software,
which is hardly surprising as it is natural for them to focus on business

18

Requirements are often already a solution

rules and user interfaces and not on the infrastructure to support it.
Because business people lack deeper technical understanding of
implementation details, their proposed solution is sometimes much
more complicated than it needs to be. It it is not uncommon for a
technical person to suggest a much simpler solution once they know
what the problem is.
In 2005, I was involved in building a J2EE-based system for Bluetooth
content distribution. It was initially used for pushing small files and
text messages to mobile phones. After a few months, the client wanted
to add on-demand content pull and distribute large animations and
video files. Enabling on-demand content pull required significant
system changes, not to mention the fact that a Bluetooth network is
not suitable for sending five-megabyte movie files to hundreds of
mobile phones at frequent intervals. As always, the client wanted it
done as soon as possible, ideally in less than two months. There was
no way that such functionality could be implemented and work
properly in time. Even if it was possible, such a hack would make
future support a nightmare.
We asked the clients to identify the problems they wanted to solve by
pulling films to the phones. It turned out that they had an opportunity
to sell the system to an art fair, where the visitors would use mobile
phones to view a video tour. From this perspective, the video tour
had absolutely nothing in common with the original solution, except
the idea of software running on mobile phones. We decided to build
a small stand-alone application for the mobile devices, which would
read all the files from an MMC memory card and would not
communicate with any servers at all. This was done in time for the
art fair and did not break the architecture of the server.
This example demonstrates how important it is to get technical people
involved in the specifications process and to share information about
the problems that we are trying to solve, not just the proposed solutions. Traditional ways of defining requirements and specifications
only harness the knowledge of a selected few individuals, and don't
really use all the brains on the team.

19

The next bottleneck in software projects

Both traditional and agile development processes expect clients and

business experts to specify what the system should do and rely on
them to get it right. I do not agree with this. Business experts and
clients should definitely be in the driving seat of projects, but I think
that much better results can be achieved if developers and testers also
have a say in what is to be built.

Cognitive diversity is very important

Getting different people involved in the process of specifying the
system is also very important from the aspect of cognitive diversity.
When everyone in the group approaches a problem from a similar
viewpoint, ideas are reinforced by a kind of echo effect and it becomes
really hard to see blind spots. People in homogenous groups often
tend to make decisions to minimise conflicts and reach consensus
without really challenging or analysing any of the ideas put to
discussion. In psychology, this effect is known as groupthink.7
Groupthink is a serious problem for decision making, not limited to
the software world. In The Wisdom of Crowds[6], James Surowiecki
blamed major US foreign policy mistakes, such as the Bay of Pigs
fiasco and the failure to anticipate the attack on Pearl Harbor, on
overly enthusiastic plans as a result of Groupthink. Instead of making
people wiser, Surowiecki claims that being in homogenous groups
can make people dumber (see [6] pp. 341).
Small homogenous groups also exhibit the effects of peer pressure
and conformity. Surowiecki quoted an experiment by Solomon Asch,
in which he showed groups of people three lines and asked them to
identify which line was the same length as a line on a card. Asch lined
up people in the group and showed them a range of cards in sequence,
asking them to give their answer out loud in the order they were lined
up. Unknown to the last person in the line, all the other subjects had
been actually told up front which answer to give. For the first few
cards, they gave the correct answer and then deliberately started
selecting incorrect lines. The last person then often started scrutinising
7

http://en.wikipedia.org/wiki/Groupthink

20

Breaking The Spirit of Kansas

the picture more, moving around to measure the lines again. About
70% of the subjects changed their answer at least once and one third
of the subjects went along with the group at least half the time. Asch
repeated the experiment with at least one of the collaborators selecting
the correct answer. This immediately encouraged the experiment
subjects to say what they really thought and the rate of conformity
plummeted. This experiment demonstrates the effects of peer pressure and how people are generally reluctant to state their opinion if
it is contrary to all the other opinions in the room. It also demonstrates that the situation changes dramatically even when a single
different opinion exists.
Surowiecki also states that even small groups of five to ten people can
exhibit what he called the wisdom of the crowds, reaching a state
where the group together is smarter than any individual in the group.
Cognitive diversity and independence of opinion are key factors in
achieving this. People should think about the problem from different
perspectives and use different approaches and heuristics. They should
also be free to offer their own judgments and knowledge rather than
just repeating what other people put forward. By getting different
people involved in the specification process, we can get the benefits
of this effect and produce better specifications.

Breaking The Spirit of Kansas

On 23 February 2008 a B-2 stealth bomber named The Spirit of Kansas
took off from from Andersen Air Force Base on the island of Guam.
By that time, The Spirit of Kansas had more than 5000 flight hours
and B-2 bombers were considered highly successful, having taken
part in various military campaigns since 1999. Shortly after take-off,
the plane started spinning and crashed inexplicably. The aircraft was
completely destroyed, with an estimated $1.4 billions' worth of
damage. The two pilots ejected in time and survived the crash with
burns. On June 6th 2008, CNN reported8 that the bomber crashed
because of an issue with control instruments, caused by specific
8

http://edition.cnn.com/2008/US/06/06/crash.ap/index.html

21

The next bottleneck in software projects

climate conditions over the island of Guam. Moisture caused three

out of the plane's 24 sensors to receive distorted data, causing the
flight-control system to send an erroneous correction command. The
pilots and crew followed the right procedures and CNN quoted Major
General Floyd L. Carpenter, who headed the accident investigation
board, as saying that the aircraft actually performed as it was
designed. In other words, all the systems were functioning normally.
Carpenter also said that some pilots and maintenance technicians
had known about these specific climate issues for at least two years
but had failed to communicate them. One of the conclusions of the
investigation was that the human factor of failure to communicate
critical information was one of the causes of the incident.
Although I was never involved in software crashes of this financial
proportion, it was not uncommon to find that someone in the
organisation had knowledge that could prevent a problem from
happening in the first place, but they failed to communicate it. Facilitating communication throughout the project is a key factor for
preventing major failures, but the traditional project model does not
encourage this. Testers in particular may be aware of some problems
that need to be addressed in development, but they do not get involved
in the effort until very late and there is nothing that facilitates the
extraction of such key knowledge from them with traditional
specifications or requirements. This is not a problem specific to
software projects, and lots of organisations struggle with extracting
knowledge from their own employees. Lew Platt, a former CEO of
Hewlett-Packard, is famously quoted for saying that If HP knew
what HP knows, we would be three times more profitable.9

Stuff to remember
The traditional specifications and requirements processes
now established in the software industry are inadequate
and essentially flawed.
Specifications do not contain enough information for
effective development or testing, they are prone to
9

See A Measurable Proposal by Tom Davenport[7].

22

Breaking The Spirit of Kansas

ambiguity and they only use the knowledge of a selected

few individuals.
Things get lost in translation between customers' original
problems and development.
Teams that do use agile development practices are
affected by this problem less than those that do not use
agile development practices, but they still suffer from it.
Unless everyone involved understands the business
goals, there is a high risk of missing the target.
Developers and testers may skip or misinterpret parts of
the specifications, and there is no easy way to check
which parts are actually affected.
Requirements are often already a proposed solution to
the problem, and do not explain why something is
required.
Gaps and inconsistencies in traditional specifications
get discovered only when developers start writing code.
Change requests may introduce inconsistencies into the
software, and there is no easy way to spot problematic
rules and affected parts of the system.
Specification documents often do not reflect the true
state of implemented software.
Customers and business analysts are expected to get
the requirements and specifications right and the
knowledge of developers and testers is not taken into
account.
The cumulative effects of small misunderstandings cause
huge problems, but there is nothing to help us discover
and resolve these issues before implementation.
Matching what clients expect is essentially a communication problem, not a technical one.
There are some things that are so obvious to the
customer that she will never tell you them unless you
ask.

23

24

Chapter 2

Finding ways to
communicate better
In order to bridge the communication gap, we need to work on
bringing business people and implementation teams together rather
than separating them with formal processes and intermediaries. Ron
Jeffries said, during his session on the natural laws of software
development at Agile 2008, that the most important information in
a requirements document are not the requirements, but the phone
number of the person who wrote it. Instead of handing down
incomplete abstract requirements, we should focus on facilitating the
flow of information and better communication between all team
members. Then people can work out for themselves whether the
information is complete and correct and ensure that they understand
each other.

Challenging requirements
Since requirements, however clearly expressed, may contain gaps and
inconsistencies, how do we fight against this problem before development rather than discovering it later? How do we ensure that
requirements, regardless of their form and whether or not they are
built up incrementally before every iteration, are complete and
correct? Donald Gause and Gerald Weinberg wrote in Exploring
Requirements[3] that the most effective way of checking requirements
is to use test cases very much like those for testing a completed system.
They suggested using a black-box testing approach during the
requirements phase because the design solution at this point still does
not exist, making it the perfect black box. This idea might sound
strange at first and it definitely takes a while to grasp. In essence, the
idea is to work out how a system would be tested as a way to check

25

Finding ways to communicate better

whether the requirements give us enough information to build the

system in the first place. We should not dive straight away into how
to implement something, but rather think about how the finished
system will be used and then double-check the requirements armed
with this knowledge. Black-box test scripts are by nature very precise
and contain clearly defined steps and values, unlike traditional
requirements which are generally a lot more abstract.
On the timeline of a software project, requirements are at the very
beginning and black-box tests are traditionally at the end. All the
things in between make it hard to see a very subtle relationship
between these two concepts. They effectively talk about the same
thing: how the system will be used once it is developed. In fact, because
tests are very precise, they offer a lot less chance for misunderstanding
than abstract requirements. This makes tests theoretically an option
for replacing requirements altogether. Robert C. Martin and Grigori
Melnik even consider them the same thing in their equivalence
hypothesis[8]:
As formality increases, tests and requirements become
indistinguishable. At the limit, tests and requirements are
equivalent.

We really communicate with

examples
In order to identify requirements, business analysts often work
through a number of realistic examples with the customers, such as
existing report forms or work processes. These examples are then
translated to abstract requirements (equivalent to the the first step
of the telephone game). An interesting thing about examples is that
they pop up several times later in the process as well. Abstract
requirements and specifications leave a lot of space for ambiguity and
misunderstanding. In order to verify or reject ideas about the
requirements, developers often resort to examples and try to put
things in a more concrete perspective when talking about edge cases

26

We really communicate with examples

to business experts or customers. Concrete realistic examples give us

a much better way to explain how things really work than requirements. Examples are simply a very effective communication technique
and we use them all the time; this comes so naturally that we are often
not aware of it. Test scripts that are produced to verify the system are
also examples of how the system behaves. They capture a very concrete
workflow, with clearly defined inputs and expected outputs.
The participants of the second Agile Alliance Functional Testing
Tools workshop, held in August 2008 in Toronto, came up with a
diagram1 explaining the relationship between examples, requirements
and tests shown in Figure 2.1. Examples, requirements and tests are
essentially tied together in a loop.

Figure 2.1. Relationship between examples, tests and

requirements

This close link between requirements, tests and examples signals that
they all effectively deal with related concepts. The problem is that
every time examples show up on the timeline of a software project,
people have to re-invent them. Customers and business analysts use
one set of examples to discuss the system requirements. Developers
come up with their own examples to talk to business analysts. Testers
come up with their own examples to write test scripts. Because of the
effects illustrated by the telephone game, these examples might not
1

Many thanks to Jennitta Andrea for pointing this out to me

27

Finding ways to communicate better

describe the same things (and they often do not). Examples that
developers invent are based on their understanding. Test scripts are
derived from what testers think about the system.
Going back to the equivalence hypothesis, tests and requirements can
be the same. Requirements are often driven from examples, and
examples also end up as tests. With enough examples, we can build
a full description of the future system. We can then discuss these
examples to make sure that everyone understands them correctly. By
formalising examples we can get rigorous requirements for the system
and a good set of tests. If we use the same examples throughout the
project, from discussions with customers and domain experts to
testing, then developers or testers do not have to come up with their
own examples in isolation. By consistently using the same set of
examples we can eliminate the effects of the telephone game.

Working together, we find better

solutions
Here is an example that shows how communication across project
roles can lead to a better solution. During the Crevasse of Doom2
presentation at the QCon conference in London in March 2007, Dan
North talked about a project in the early days of Java, when printing
documents was extremely hard to implement properly. The developers
told the business people that all the requirements could be easily
implemented, apart from those related to printing because printing
APIs were not mature enough. The business people refused to give
in because printing was embedded in the core of their business.
While discussing this further and talking about how exactly printing
was embedded in the core of the business, the developers discovered
that the users entered data in a screen, printed out the results of
calculations, opened a different screen and entered these results,
processed and printed results again, entered them in the third screen
and so on. After seeing this, the developers suggested automatically
2

http://www.infoq.com/presentations/Fowler-North-Crevasse-of-Doom

28

Communicating intent

transferring the information between screens, invoking the amazed

response You can do that?. Although this example is rather
simplistic, it illustrates a very important fact: neither group could
solve the problem on their own, but together they found a great
solution. The business users were so used to working around a technical difficulty that they really thought of printing as something
embedded in the process. The developers did not really understand
what was going on or why printing was needed.
If we help each other understand the goal better, then we will come
up with better solutions. This is why building a shared understanding
of the goal is one of the key practices in software development.

Communicating intent
The communication of goals was at the core of Prussian military
tactics as a response to the dangers posed by Napoleon's invincible
army. The Prussian leaders figured out that they did not have a single
person capable of defeating Napoleon's genius, so they focused on
allowing the individual commanders and their troops to act collectively to better effect. They made sure they told the commanders why
something needed to be done and they put this information into the
perspective of overall goals, rather than just passing down a list of
imperative commands. The resulting military doctrine was called
Auftragstaktik, or mission-type tactics, and it was key to the great
successes of Prussian and later German armies. Today it survives as
Mission Command in the US Army.
A key lesson to take from this and apply in software development is
that understanding business reasons behind technical requests is
crucial for building a shared understanding of the domain. Examples
in the previous section and the section Requirements are often already
a solution on page 18 demonstrate this. Original intent is one of the
most important things that a customer or a business analyst should
pass on to the developers and testers. At the same time, in my experience, passing on this information is one of the most underestimated
and neglected practices in software development. This is just as much

29

Finding ways to communicate better

a problem in other areas. Going back to the example with US Army

orders from the section Imperative requirements are very easy to
misunderstand on page 8, effectively communicating the intent was
one of the biggest problems that researchers identified. Gary Klein
cites research by William Crain,3 who found that only 19% of
commanders' intent statements had anything to say about the purpose
of the mission and that communication of intent was mediocre.
When domain experts reviewed the statements, the average score for
effectiveness was below the average, closer to very ineffective.
I was recently involved in a project where the clients said that they
wanted real-time reporting on transactions in a highly distributed
system. The system was being developed for batch reporting and
introducing real-time reporting required a huge redesign, possibly
even a major change to the infrastructure. Instead of challenging this
requirement, the CTO of the company started to design a solution
on the spot, immersing himself in a tough technical problem which,
for him, was a pleasure to play with. I asked what the clients wanted
to achieve with real-time reports and we came to the conclusion that
they wanted to present up-to-date information to their partners during
the working day, not just overnight. A delay of one hour was perfectly
acceptable for this, without the need to redesign the system from
scratch or invest a lot of money in real-time infrastructure, especially
since real-time reporting on live transactional data would have seriously impacted the performance of the transaction processing service.
In this case, the issue was with the meaning of word real-time. For
the customers, it meant updated several times per day but for
programmers it meant millisecond correctness.
The moral of this story is that requirements should not really be taken
at face value, especially if they come in the form of a solution that
does not explain what the intent is (see the section Requirements are
often already a solution on page 18). Teams often consider requirements as something carved in stone that has to match exactly what
the customers demand. In fact, customers will often accept a different

See page 225 of [4] and http://handle.dtic.mil/100.2/ADA225436

30

Agile acceptance testing in a nutshell

solution if it solves the problem, especially if it is a lot cheaper and

easier to implement.
Understanding why something is required, instead of just blindly
following instructions, is definitely one of the key factors for
improving the chances of success in a software project. This is why it
is crucial to communicate the reasons behind technical requests and
business decisions. Don't be afraid to ask why if you do not understand a requirement or if you find it strange.

Agile acceptance testing in a

nutshell
We can bridge the communication gap by communicating intent,
getting different roles involved in nailing down the requirements and
exploiting the relationships between tests, examples and requirements.
We can use realistic examples consistently throughout the project to
avoid the translation and minimise the effect of the telephone game,
saving information from falling through the communication gaps.
In order to ensure that the initial set of examples is good enough for
development and testing, programmers and testers have to get
involved in defining them. The discussion about these examples builds
up a shared understanding of the domain, flushing out small
communication problems straight from the start.
This is essentially what agile acceptance testing is about. Over the
years, it has evolved into a relatively simple and elegant practice.
(Please do not confuse simple and elegant with easy. Implementing
agile acceptance testing can be a real organisational challenge and
may require a lot of effort. But it will pay off in a big way at the end.)
It revolves around the following five ideas:
1. Use real-world examples to build a shared understanding of the
domain.
2. Select a set of these examples to be a specification and an acceptance
test suite.

31

Finding ways to communicate better

3. Automate verification of the acceptance tests.

4. Focus the software development effort on the acceptance tests.
5. Use the set of acceptance tests to facilitate discussion about future
change requests, effectively going back to step 1 for a new cycle of
development.

Use real-world examples to build a shared

understanding of the domain
Instead of abstract requirements that can easily be misinterpreted,
we should talk about what the system does or should do in real-world
situations. Abstract definitions can obscure our view and leave many
ambiguities. Real-world examples do not suffer from this. Discussing
realistic situations helps flush out assumed business rules and points
us to the true business rules. Ideally, we want to identify representative
cases that show all the differences in possibilities and then discuss
these cases and how the system should behave. Writing the examples
down on a whiteboard, wall, word document or a piece of paper will
help discover gaps and identify additional scenarios and examples
that we have to discuss. Working with real-world examples helps us
communicate better because people will be able to relate to them more
easily. It is also easier to spot inconsistencies between realistic
examples. Developers, business people and testers all need to participate in the discussion about examples. Developers learn about the
domain and get a solid foundation for implementation. Testers obtain
the knowledge they need firsthand and they can influence the development by suggesting important cases for discussion. I explain this
step in greater detail in Chapter 3 and Chapter 4.

Select a set of these examples to be a

specification and an acceptance test suite
Once we have identified enough examples for everyone to be
comfortable with starting the implementation, we can actually use
the examples as the acceptance criteria for the current iteration of

32

Automate verification of the acceptance tests

development. A selected set of examples can become the target for

development effectively the solution specification. Once the system
does exactly what we discussed in all of the cases, it is ready to be
shipped. Verifying all the cases becomes the final check-point before
a release. This step is explained in greater detail in Chapter 5.

Automate verification of the acceptance

tests
If we are moving away from the target, I'd like to know this sooner
rather than later. So we want to be able to check the examples often,
to make sure that the code is going towards the target. The biggest
expense in software development today is the time of the implementation team members (again, I'm using this as a name for all participants in the development process, not just developers), so verifying
the examples must not take a lot of time. In addition, automated tests
provide quick feedback and make the rest of our work more efficient.
This is why we need to automate the verifications. When all the
examples are automated, we can check whether the code does what
we expected with a touch of a button. This step is explained in greater
detail in Chapter 5.

Focus the software development effort on

the acceptance tests
At the end of the project all the examples should work. Because we
ideally have a complete set of examples, if all the examples work there
is nothing else that should be developed. So we can actually focus the
development on satisfying the examples. After the examples are
automated, we can verify the code quickly and often. Develop a bit
of code to satisfy an example, run the verification and check whether
we are really doing it correctly, then move on to the next example.
This helps to avoid just-in-case code, which is what happens when
developers have to think about all the edge cases to satisfy an abstract
requirement. There are no abstract requirements, and we have

33

Finding ways to communicate better

examples to describe all the edge cases. When all the examples are
implemented so that the system works as they describe, the job is
done. To developers this might sound similar to unit testing, because
it essentially tries to do the same thing on a higher level (but acceptance tests are not a replacement for unit tests). I talk about this step
in more detail in Chapter 6.

Use the acceptance tests to facilitate

future change
Once the development is done, change requests will start coming in.
We can use the set of acceptance tests for previous phases of development as a relevant authoritative document on the system. We use
them to discuss change requests and quickly discover conflicting rules
and changes. This is explained in more detail in Chapter 7.

And repeat
These steps are continuously repeated throughout the project to
clarify, specify, implement and verify small parts of the project iteratively and incrementally. In Chapter 8 I explain how this fits into the
overall development process in more detail.

So what does testing have to do with

this?
On the first page of the introduction I said that agile acceptance
testing is a practice that improves communication in the team. I also
talked about focusing development and producing better specifications. This often raises the question of what all this has to do with
testing. In fact, the name agile acceptance testing is a bit misleading
and one of the main causes for so many misconceptions about this
practice. The name comes from the fact that this practice evolved
from test-driven development (TDD). Don't confuse it with customer

34

So what does testing have to do with this?

(or user) acceptance testing, which is a practice that was established

a long time before agile acceptance testing. I explain the differences
between these two later.
In test-driven development, developers write tests to specify what a
unit of code should do, then implement the code unit to satisfy the
requirement. These tests are called unit tests, as they focus on small
code units. The benefits of this approach are a better focus on what
the system should do and having a clearly defined target for development. The team can collectively own the code and verify it, so this
practice facilitates better team work. Unit tests also hold the system
together during change, allowing it to be more flexible. If a change
unintentionally breaks some existing functionality, the relevant unit
test will fail and alert developers about a problem.
The practice of unit testing was so useful that it was only logical to
ask whether the same could be applied to business rules and drive
whole phases of projects rather than just code units. Unit tests focus
on code so they are completely in the domain of software developers.
Business rules, on the other hand, should not be defined by developers.
They have to be defined by domain experts and customers. But business domain experts rarely understand programming languages, so
using the same tools to drive implementation of business rules and
code units typically fails. Developers could write business rule tests
with unit test tools, but these tests would only reflect what developers
understand and they would still be affected by all the communication
problems presented in Chapter 1. In order to get the best specifications, customers and implementation teams have to work together.
There is simply no way for a business person to verify that a
developer's test describes the end goal correctly with unit test tools.
So, in theory, the ideas behind unit testing could be applied to business
rules as well, but in practice the tests are impossible to communicate.
To solve the problem, better ways for specifying and automating tests
for business rules were needed that could also be used to communicate
with business people. Such tools would have to focus on capturing
the customers' view of what the system should do when it is finished,
effectively the functional acceptance criteria for the project. Then we

35

Finding ways to communicate better

could apply unit testing ideas by developing code to satisfy tests and
running these tests to verify that the code is on target, then repeating
the process until all tests go green. This is where the name agile
acceptance testing comes from. This practice started by expanding
the unit testing ideas to business rules, effectively specifying the
acceptance criteria in a form that could be executed as tests on the
code.
Going back to the conclusion of the previous chapter, the problem is
essentially a communication issue, not a technical one. The biggest
obstacle for any such effort is communication, so easy communication
and collaboration with business people have to be part of any viable
solution. Out of the effort to solve the problem have come tools like
FIT, FitNesse and so on. The ideas applied in these tools solved the
communication problem so nicely that the practice of agile acceptance
testing evolved into a great way to build a shared understanding of
the domain and enable all project participants to speak the same
language. Testing, in any possible meaning of this word, becomes
relatively unimportant because the greatest benefit is improved
communication.
Today, the name of this practice itself has become a major obstacle
to its adoption. The chief business analyst of a company I recently
worked with just rejected getting involved in the practice, with the
explanation I do not write tests. The word testing unfortunately
bears a negative connotation in the software development world. In
all the companies I worked for, testers were among the least well-paid
employees, right down there with the support engineers. Starting a
discussion on testing somehow seems to give business people the
green light to tune out. It is like a signal that the interesting part of
the meeting is over and that they can start playing Sudoku or thinking
about more important things. After all, testing is not something that
they do.
The practice of agile acceptance testing, especially as a way to improve
communication between team members and build a shared understanding of the domain, relies on the participation of business people.
They are the ones that need to pass their domain knowledge on to

36

So what does testing have to do with this?

programmers and testers. They are also the ones that need to make
decisions about edge cases and answer tough questions about business
rules. So they very much have to do tests.
To add a further complication, there is also user acceptance testing,
which sounds very similar to agile acceptance testing. User acceptance
testing (UAT) is a phase in a software project where clients sign off
a deliverable and accept it as complete. It can involve testing by endusers, verification from stakeholders that they are happy with the
product and a whole range of other testing activities. Although you
can verify the software during UAT using a successful run through
an acceptance testing suite as one of the criteria, agile acceptance
testing has very little to do with user acceptance testing. In fact, they
are on completely opposite sides of a software project. User acceptance
testing happens after development and it is often the final confirmation before money changes hands. It is performed by the clients or
third-party agents. Agile acceptance testing happens before and during
development and it is performed by the implementation team.
Some people call this practice acceptance test-driven development to
signify that acceptance tests are used to drive the development, not
just to verify the deliveries at the end. Others try to reduce the
confusion between user acceptance testing and agile acceptance testing
by avoiding the use of the word acceptance and calling the practice
functional test-driven development. This also signals the difference
between code-oriented unit tests which are traditionally linked with
TDD and functional (acceptance) tests. Three or four years ago, the
name storytest-driven development was used to describe this practice,
emphasising the fact that acceptance tests are related to user stories
and not code, but it did not really catch on. Another variant is testdriven requirements, explaining that acceptance tests actually deal
with requirements more than with development. Brian Marick calls
these tests business-facing to emphasise that business people should
be concerned with them.4

http://www.exampler.com/old-blog/2003/08/21/

37

Finding ways to communicate better

Better names
Dan North suggests using the word behaviour instead of test,5 as
a way to clear up a lot of misunderstandings. Instead of test-driven
development, he talks about behaviour-driven development to avoid
the negative connotation of testing. This trick has solved the problem
of keeping business people awake quite a few times for me as well.
Behaviour-driven development (BDD) is just a variant of agile
acceptance testing, in my opinion. Some people will disagree, pointing
out the differences in tools and format of test scripts. For me, the
underlying principles are the same and BDD tools are just another
way to automate tests. BDD also promotes a specific approach to
implementation,6 but this is not really important for the topic of this
book. Again, I consider tools and tests to be of less importance than
communication and building a shared understanding.
A name that has become more popular recently is example-driven
development, with tests being called specification by example. This
reflects the fact that we are using concrete real-world examples to
produce the specifications instead of abstract requirements. I also
like executable specifications as a name instead of acceptance tests,
because it truly describes the nature of what we are building. Agile
acceptance tests are specifications for development in a form that can
be verified by executing them directly against the code.
People have tried to rename agile acceptance testing several times,
but this name has somehow stuck. Most still use this name and this
is why I decided to keep it here. I mention all the alternative names
here because they are interesting attempts to remove the word test
from the vocabulary to reduce the ambiguity and misunderstanding
that come from it. If your business people don't want to participate
because testing is beneath them, try to present the same thing but
with a different name.

5
6

http://dannorth.net/introducing-bdd
See http://behaviour-driven.org/ for more information

38

Better names

In any case, I want to point out that there are a lot of different names
and ideas emerging at the moment, but they are all effectively different
versions of the same underlying practice. This book is about the
underlying values, practices and principles that all these names and
ideas share.

Stuff to remember
Realistic examples are a great way to communicate, and
we use them often without even thinking about it.
Requirements, tests and examples all talk about the
same thing how a system will behave once it is
delivered.
We can use examples consistently throughout the project
to avoid the effects of the telephone game.
Building a shared understanding of the problem is one
of the key practices in software development.
Business intent is one of the most important things that
a customer or a business analyst should pass on to the
developers and testers.
Cross-functional teams create much better specifications
and requirements than business people in isolation.
Agile acceptance testing uses these ideas to solve
communication problems on software projects.
The name agile acceptance testing is misleading but has
been generally adopted.
Agile acceptance testing is very different from user
acceptance testing, and in general it is not about testing
at all. It is about improving communication and building
a shared understanding of the domain.
Implementing agile acceptance testing can be a real
organisational challenge.
Agile acceptance testing in a nutshell revolves around these
five principles:

39

Finding ways to communicate better

1. Use real-world examples to build a shared understanding

of the domain.
2. Select a set of these examples to be a specification and
an acceptance test suite.
3. Automate the verification of acceptance tests.
4. Focus the software development effort on the acceptance
tests.
5. Use the set of acceptance tests to facilitate discussion
about future change requests, effectively going back to
step 1 for a new cycle of development.

40

Part II. Building and Maintaining

a Shared Understanding
Using examples as specifications consistently throughout the project helps
to avoid the effects of the telephone game and enables us to truly build a
shared understanding of the system. In this part, we look into the roles
of examples from specifications through development and testing to
becoming a live documentation on the system. We also look into practices
that enable us to get the most out of those examples.

Chapter 3

Specifying with
examples
The first stage of agile acceptance testing is to make sure that we all
know what we are talking about and more importantly to agree on
this. Before a phase of development, be it an iteration, a mini-project
or simply a chunk of software whose time has come, we specify what
we expect out of it. We specify this in the form of real world examples,
not with abstract requirements. The examples demonstrate how the
system should act and how it should help users do their jobs. These
examples are created by the whole implementation team, not by a
single domain expert as in the traditional model. We use the examples
to discuss the domain and make sure that there are no misunderstandings.

How do you brush your teeth?

Traditional requirements, even when written without functional gaps,
leave a lot of room for misinterpretation because of hidden assumptions. People coming from different backgrounds see the same thing
differently. They may not have the same interpretation when they
read or hear something, as explained in the section Are obvious things
really obvious? on page 9.
As a very effective demonstration of this, Mickey Phoenix asked people
to describe how they brush their teeth at the Agile 2008 Conference
in Toronto in August 2008, during his session on domain-specific
testing languages.1 Having proposed a very simple description such
as put toothpaste on the toothbrush, open mouth, brush teeth..., he
pointed out that the results of performing the first step might actually
1

http://www.solutionsiq.com/agile2008/agile-2008-domain.php

43

Specifying with examples

look similar to the picture in Figure 3.1 to someone who has never
seen a toothbrush before.
If we don't share the same understanding of the business domain,
even the simplest of explanations can be ambiguous and misinterpreted. Software developers are technical experts and they know how
to write code, but they often have no real experience of the business
domain and their assumptions can be substantially different from
the assumptions of business people. Unless we build a shared understanding and flush out assumptions, developers will often end up
putting the toothpaste on the wrong side of the brush.

Figure 3.1. Put the toothpaste on the toothbrush

A live demonstration beats a list of instructions at any time show

people an example and they will learn a lot more and a lot faster. I
noticed a short TV clip on one of the children's TV channels, which
teaches viewers how to brush their teeth. Instead of just talking about
it, the presenter demonstrated all the steps as he was explaining what
kids should do. For those among you who will challenge this example
for being overly simplistic and applicable only to children, here is
something more relevant to adults in business. Try to explain how to
tie a Windsor necktie knot only using words. I challenge you to explain
it well enough so that someone can tie a correct Windsor knot based
on your description. If you type Windsor knot into the search field
in Google, you will get loads of sites explaining how to tie the knot

44

A practical example

with images and even some videos I have not found a single site in
the first few result pages that does it only with words.

A practical example
One of the best ways to ensure that people understand each other is
to demonstrate various differences in possibilities with realistic
examples. In the poker example in the section A small misunderstanding can cost a lot of money on page 14, if there had been a discussion
between developers and business people including the case where a
cent is rounded to a penny, the business analysts would have spotted
the issue straight away. Abstract requirements and specifications are
not a good tool for communication. Real-life examples are much
better.
Several years ago I was involved in building an affiliate advertising
system. Affiliate advertising systems, for those of you who have slept
through the web advertising revolution, connect web sites that want
to advertise something and web sites that offer advertising space. The
company that owns the web site with free space is called the affiliate.
The company that sells products or services and wants to advertise
pays a commission for the customers that come from the affiliate site
by clicking on an advert. Today, the bulk of such commission is paid
as a percentage of sales, but back then it was also paid on the number
of clicks.
Our client wanted to build their own affiliate management system
rather than join an existing ad exchange. For the sake of simplicity,
let's say that their initial request was to pay affiliates 2 pounds, dollars
or euros (whichever your preferred currency is) for 1000 clicks. As a
requirement, pay 2 pounds per 1000 clicks does not seem strange at
all in fact it is very much like any other normal requirement.
However, although it seems precise, it leaves a lot of room for speculation. What happens if the affiliate has only 500 clicks on a particular
day? Do we pay him 1 pound or do we not pay him anything at all?
As a developer, I can think about this question and give an educated
guess or opinion, but the real truth is that I should not be deciding

45

Specifying with examples

about such an issue. This is a question for the business people because
it is a manifestation of their business model. Maybe they want to pay
a pound, maybe they want to wait for people to accumulate 1000
clicks. Both options are valid from a technical perspective, but there
is only a single valid option from the business perspective. If we leave
this question to developers to decide, there is a big chance that they
will not select the same option as the customers did.
Problems like this one are not really bugs in the classical sense they
are caused by misunderstandings of business rules. The code might
be completely correct from the unit testing perspective, but still miss
the business target. This is where real-world examples come in. Instead
of an abstract requirement such as pay 2 pounds per 1000 clicks, we
need to identify interesting realistic cases and then discuss what
happens in these cases.
For a start, it is very rare for a web site to get a nice round number of
clicks. So a much more realistic case would be to have something like
7672 clicks during a day. What do we do with this number of clicks?
Do we round it down and pay 14 pounds, or do we round it to 8000
clicks and pay 16 pounds? Maybe just scale it and pay 15.34 pounds?
I like to throw in edge cases as well. Should we pay anything for 999
clicks? What happens with just one click? Discussing edge cases like
these would have pointed out the problem with 1 cent in the foreign
exchange story.
In this case, we could write the examples on a whiteboard, Excel
spreadsheet or something like that and then identify the cases.

46

A practical example

We then get the customers to give us values that they expect to pay
in these situations:

Discussing cases like these often raises more questions and reveals
other interesting examples. Because we don't pay anything for 999
clicks, it is not really fair to the affiliates to simply ignore their total
if they are missing a single click. In this case, the next question might
be should we let the the 999 clicks roll over to the next day or do we
reset the counter to zero for tomorrow? Maybe the clients are not too
bothered with small web sites, they only want to focus on big advertisers who will have several thousands of clicks per day, so discarding
this small remainder is not a big issue. Or maybe they also want to
please the small advertisers, who will rarely have more than a thousand
clicks on a day, but would appreciate being paid for the accumulated
clicks. So we can create a few more examples on the whiteboard and
discuss them:

47

Specifying with examples

If we do allow the roll-over, then the example may raise additional

questions. How many days do we allow the roll-over? What happens
in the following case?

This may raise additional questions. Is there a limit to roll-over, a

cut-off at the end of the month or something similar? Do we just let
the clicks roll over and accumulate?
Although this example is a bit simplistic, this kind of process is very
common once you start discussing realistic examples. Edge cases will
lead to more questions, more examples, discussion, more questions
and so on. The answers to all these questions depend on the customer's
business model and they are not for the implementation team to
answer, but it is their job to raise the right questions. The important
fact is that these tough questions need to get asked and resolved before
people start working on the code, not midway through implementing
a class while the customer or business analyst is away and unreachable
by phone. Just to make sure that there is no misunderstanding, I am
not talking about having this discussion at the very beginning of an
entire project. A workshop at the start of an iteration fits the before
people start working on the code criterion just fine.

Realistic examples make us think

harder
Compared to abstract specifications or requirements, realistic
examples offer a lot less chance for misunderstanding. We have to
specify exactly how we want the system to work in a particular case.

48

Realistic examples make us think harder

Because examples get written down and discussed in detail, it also

becomes easy to spot gaps, missing cases and inconsistencies that
would otherwise have been noticed only during development.
While working through examples for a financial fraud detection
system, I noticed that several of them had special cases for VIP
customers. We then went back to the examples that did not have any
special cases and asked whether the VIP status should be considered
there as well. It turned out that several of these use cases (such as the
freezing of funds) actually should not be applied to VIP customers.
Problems like these would otherwise surface in production, after an
angry call from a VIP customer. Discussing realistic examples allowed
us to identify the gap and close it before the development started. I've
heard an argument that this has nothing to do with working with
realistic examples and that a good business analyst would have identified a problem such as this. I beg to differ. In this particular situation,
there were eight people in the room among which four were domain
specialists and one was an experienced senior business analyst. We
all failed to spot the functional gap in the first round of analysis.
Working with realistic examples made it much easier to spot that
requirements were not consistent and provided a spark for the
discussion. No matter how good a single person is, several people
discussing realistic examples from different perspectives will flush
out inconsistencies better.
Realistic examples contain precise information and they ask for precise
answers. This often makes people think harder and not just brush the
question off. It is not uncommon for people to disagree on edge cases
during the discussion and actually fine-tune their business processes.
I remember a situation where we discussed selling media clips online
and asked what should happen if a customer wanted to buy a $10 clip
but only had $9.50 in his account. Although the first response was
that the sale should be refused, the CEO of the client company decided
that it made more sense to allow the purchase to go through and to
put the account into negative balance by 50 cents. This would increase
sales and the customer would have to top up the account anyway
before the next purchase. Developers and testers can sometimes
provide a completely different perspective on a business problem.

49

Specifying with examples

Although I would not consider this situation as an example of how

companies should create their business strategies, viewing things
from a different perspective can sometimes help people spot new
opportunities.

Identifying important examples

With agile processes, a feature is often implemented by incrementally
adding functionality through several iterations. Mike Scott suggests2
focusing first on a spine story which contains the essence of a feature,
and then adding details and exceptions in later iterations. Planning
projects and splitting a feature into multiple iterations are subjects
that are really beyond the scope of this book, but I do discuss them
briefly in Chapter 9. For now, I want to point out that we need to
focus on the examples that are really important for a particular phase
of development, however it is planned. We should keep the future in
mind in order not to paint ourselves into a corner, but it is important
not to over-complicate things or to get bogged down in over-analysing
all possible outcomes.
We want to gather as many representative examples as required to
show all the possibilities and all the important cases for a feature for
the current iteration. A good starting point is to take a description
of a future piece of software, such as a use case, user story or whatever
the unit of development in your planning method is, and then ask
the customers how should we verify that this piece of software is
implemented completely and correctly?. Another question I often
use is pretend that it's magic and that we already delivered this thing
how would you actually test it?. Then these examples need to be
written down, and discussed in detail.
We should then identify edge cases, negative paths and scenarios in
which things do not go according to plan and discuss them as well.
Far too often I've seen business people concerned with only the
scenarios when everything is OK, leaving the negative cases for
2

in a private e-mail

50

Identifying important examples

programmers to sort out. Again, you really do not want programmers

to make decisions about your business model. Donald Gause and
Gerald Weinberg ([3] pp. 251-2) suggest that a requirements test
(which relates to a set of examples in agile acceptance testing) must
consider the following three categories:
1. normal use
2. abnormal but reasonable use
3. abnormal and unreasonable use
Unreasonable in this case refers to something that a sane person
would not do, but it can still realistically happen. These cases are most
likely to be identified by testers because they think about breaking
the system. Gause and Weinberg point this out saying that No single
failure of requirements work leads to more lawsuits than the confident
declaration: No sane person would do that.
Team members often know about edge cases and potential problems
and even know how to resolve them, but they need to communicate
them effectively. Had the team members communicated their problems and concerns effectively, the plane described in the section
Breaking The Spirit of Kansas on page 21 might still be flying. To
avoid expensive crashes like these in our software, we should all focus
on flushing out as many of these important examples as we can.
Testers in particular are expected to suggest examples that demonstrate problems which they would typically check in later testing. In
the next chapter, I introduce the specification workshop, which
facilitates such a discussion.
If the example mentions numbers, I often throw in some numerical
edge cases just to shake things up a bit and see if something interesting
comes out. Reduce the numbers in the example by a very small
amount, suggest some very small and very large figures. They will not
necessarily lead to differences in processing, but may very well lead
you to discover something that was missed the first time. Examples
of this would be the 0.01 transaction in the foreign exchange example,
999 clicks in the web site advertising example and the media clip
purchase for $9.50. When discussing very large or small amounts, ask

51

Specifying with examples

first whether they are realistic or not. Keep in mind that there is a
difference between unreasonable and unrealistic. We don't want to
waste time discussing imaginary cases that are not important for the
system.

Differences in similar examples

Watch out for differences identified in similar or related examples.
Make sure to check whether any of the special cases should be
considered for related examples as well. More often than not, this will
help you discover gaps in the functionality. This is a lot easier to do
when the examples are written down on a whiteboard or a wall,
because you can see groups of examples together. Try to identify a
common structure in the examples and write related examples close
together in a similar form to spot such inconsistencies more easily.
If the examples differ in one or two variables, then write them as a
list or a table. Empty cells or combinations missing from the table
might point you to more interesting cases.
Consider these few examples for an online flower shop that offers
reusable free delivery vouchers to VIP clients when they make a single
purchase over $50:
Mark is a VIP customer. He puts a $50 flower arrangement in the
shopping cart and goes to checkout. On checkout, he is offered free
delivery since he is from the US.
Mark has a previous unused free delivery offer. He puts a $30 flower
arrangement in the shopping cart and goes to checkout. On
checkout, he is offered free delivery.
Mark has no previous unused free delivery offers. He puts a $30
flower arrangement in the shopping cart and goes to checkout. On
checkout, he is only offered regular delivery.
Lucy is a VIP customer from the UK. She puts a $50 flower
arrangement in the shopping cart and goes to checkout, but she is
offered only regular delivery since she is from the UK. We do not
offer free delivery outside the US. We give her a free gift voucher
instead.

52

Differences in similar examples

Tom is not a VIP customer. He puts a $50 arrangement in the

shopping cart and goes to checkout. On checkout, he is only offered
regular delivery.
The workflow steps and user names are not really important for these
rules. Instead of writing down each scenario as a separate example,
we could just have a single list that specifies something like this:

Writing things down like this makes it easier to spot some other
interesting cases that need to be discussed. Obviously, we could write
an example for a VIP customer from the UK that has only $30 worth
of flowers in the cart. But there are some specification gaps that are
less easy to spot, and the table uncovers them as well. For example,
what do we do when someone has an unused free delivery offer, but
they are a VIP customer and have more than $50 of flowers in the
shopping cart? Do we keep the free delivery offer for the next time?
What if they don't decide to use free delivery this time as well do
we allow them to keep two offers of free delivery for later or just one?
What happens if for some reason a UK customer has a free delivery
offer? How can this happen (perhaps we let someone change their
address but keep a free delivery offer), and should we disable this
option in this case? Ideally we want to flush out these important
examples while we have business experts available to discuss them
immediately.

53

Specifying with examples

Tables make it easy to spot differences

Those of you who have encountered FIT before will now surely notice
how these tables can directly map to FIT acceptance tests (I discuss
FIT in Chapter 10). At this point, I want you to disregard what you
know about FIT or any similar tool, including your likes and dislikes.
Even if you use a different tool that does not work with tables, do not
throw away the baby with the bathwater. Think of tables primarily
as a very good discussion-facilitating mechanism. Rick Mugridge
and Ward Cunningham suggest that tables provide just enough
structure to efficiently organise information, without getting in the
way and over-complicating with formalisation ([9] p. 28). Tables were
used to focus discussion and get better requirements long before agile
acceptance testing and test-driven development were invented. David
Parnas used them on the A-7 project for the US Naval Research Lab
in 1977. He wrote about his experiences in 1996[10]:
Tabular notations are of great help in situations like this. One
first determines the structure of the table, making sure that the
headers cover all possible cases, then turns one's attention to
completing the individual entries in the table. The task may
extend over weeks or months; the use of the tabular format
helps to make sure that no cases get forgotten.
Regardless of how you write them down, working through a set of
suggested examples will result in identifying new examples and
interesting cases that were not covered by the initial set. This is
perfectly normal and it is actually what we want. Be very suspicious
if nobody identifies additional examples this most likely indicates
that people are not really participating.

Dealing with processing workflows

If the user story (or use case) contains some processing workflow, I
like to draw the workflow on the wall or whiteboard and keep it there
during the whole discussion. Decision points and processes in the
workflow need to be discussed on their own, but keeping an eye on

54

Dealing with processing workflows

the big picture also helps to identify more important examples to

discuss.
For example, Figure 3.2 is a picture of a bet placement workflow for
an online betting system. Business people see this whole workflow as
an integral process, but most of the boxes in the picture need to be
discussed individually as they represent individual business rules.

Figure 3.2. Bet placement workflow

When we decompose a complex specification into a number of simpler

specifications, we can focus better on each individual rule. For each
of the decision boxes, there will be at least two or three important
cases to cover. If we were to discuss everything as a single scenario,
the number of important cases would be the product of the numbers

55

Specifying with examples

of important cases at each decision point. In this example, this would

be around 50 assuming that each step has two or three different
important examples. If we discuss business rules individually, then
the total number of cases is the sum of all important examples. In this
example, this would be something between 10 and 15. These numbers
apply only if we do not identify any important edge cases during the
discussion, which is most unlikely. In a more realistic scenario, many
edge cases would be identified and the difference between the numbers
would be much larger. The combinatorial explosion effect makes all
the many possibilities of a complex workflow really hard to describe
and comprehend precisely. By focusing on individual rules within
the workflow, we can have a much more meaningful discussion.
Depending on how you structure the deliverables of your project, it
is not uncommon to implement individual steps in different phases
or to different standards,3 which means that not all examples need
to be discussed at once. In this particular project, the automatic topup of accounts could easily be put into the next phase of development
and risk controls might not be implemented completely from the
start. Drawing a big-picture diagram like this will help you see which
parts may be postponed for later or implemented just partially.
Seeing the big picture is also important from the aspect of identifying
functionality gaps. If three out of four decision points consider
whether someone is a VIP customer or not, just ask whether the fourth
decision point should be concerned with this as well. Maybe yes,
maybe no, but let's make sure.
I do not advocate using any special or precise diagramming technique
to draw the process. As long as everyone in the room understands
what the diagram shows, any sort of drawing is fine. You can use
UML, algorithm diagrams or just write steps in English and connect
them with arrows. Don't use the discussion about examples as a chance
to practice your UML skills if other people in the room do not
understand UML.

http://gojko.net/2007/12/04/waterfall-trap/

56

Working with business workflows

Working with business workflows

Sometimes a set of examples only makes sense in the context of a
workflow and it is really hard to look at them just as a collection of
rules. An example might be the sequence of steps in a wizard helper
or a security alert escalation procedure. When dealing with such
workflow examples, try to organise them in a way that clearly identifies
what the preconditions are, what the workflow steps are and what are
the expected outcomes of the workflow. Thinking in terms of these
three aspects, sequenced in this order, will help you focus the discussion on a single rule and identify important issues around it. The
behaviour-driven development scenario template is a useful way to
write workflow examples down. It consists of three parts:
1. Given preconditions
2. When actions or triggers
3. Then consequences
For example, a simple scenario describing credit card transaction
approval would be:
1. Given the customer has a balance of $0, and the customer has a
deposit transaction for $10 and the transaction is in the authorisation state
2. When an AUTH-8440 message comes from the bank
3. Then the transaction is accepted and the customer has a balance
of $10
I find that tables are more useful for expressing specifications, state
machine transitions and calculation-based rules. The expressiveness
of natural language and the sequential style of writing help to get a
better understanding of the process for workflows. Again, this format
may seem familiar to people who use BDD tools, but this is irrelevant.
Even if you do not plan to use a BDD-style tool, try to write business
workflow examples in this format because it will enable you to have
a more meaningful discussion about them. Each flow example should

57

Specifying with examples

focus on a single rule, so that the discussion can focus on this as well.
If larger flows are written as a single example, it's often hard to keep
track of the context and it becomes easy for ambiguity and hidden
complexities to pass undetected.

Stuff to remember
Instead of abstract requirements, use realistic examples
to demonstrate differences in possibilities.
The whole team should be involved in working through
edge cases and resolving ambiguities.
Discussing realistic examples makes people think harder
and they are less likely to just brush questions off.
Watch out for small differences in examples, as they
might indicate that a business rule is not directly
expressed.
Write specifications down as tables to make it easier to
grasp the content and spot gaps and inconsistencies.
With processing workflows, discuss rules for each
decision point separately.
For genuine workflows, write down examples so that
preconditions, processing steps and verifications are
clearly separated.

58

Chapter 4

Specification
workshops
Just writing the specifications or requirements as realistic examples
instead of abstract sentences is not enough. To get the most out of
realistic examples, we need to put them to an open discussion and
give everyone a chance to review them and ensure that we all understand the same thing.
A customer or a product owner would typically be concerned mostly
with the happy path, such as paying two pounds for every thousand
clicks. Developers tend to focus much more on edge cases and
alternative scenarios, so a developer might suggest an example with
999 clicks for discussion. Testers often think about how to break or
cheat the system, so a tester might suggest the case where all 1000
clicks come from the same IP address in a time frame of five seconds.
Do we still pay the affiliate in this case, or do we ignore it? A developer
might then again ask what if this same affiliate had 3000 more clicks
on the day do we pay six pounds for these further clicks or do we
ignore the affiliate completely?
All these people have different views of the system and ultimately
different mindsets. This is why examples have to be put to discussion
and analysed by the whole team. To facilitate this process, I like to
gather customers, business domain experts, developers and testers
around a whiteboard or a wall that we can write on and discuss
interesting examples. If the team is small, then the whole team gets
involved in this. With larger teams this might be a challenge because
it is hard to keep ten or more people focused on a single thing. For
larger teams, get at least one person from each of these groups in the
room. At the start of an iteration, this turns into an intensive, handson problem and domain exploration exercise, driven by examples. I

59

Specification workshops

call this exercise the specification workshop. The workshop is focused

on a particular piece of future development, a single iteration, and
in it we try to iron out as many problems and identify and close as
many functional gaps and ambiguities as possible.
In addition to domain specialists, testers and developers, it is ideal
to have a project stakeholder present during each workshop as well.
In fact, one of the most common mistakes with workshops is not
including stakeholders. The goal of the exercise is to nail down scope.
In order to do this, you need someone who has the ultimate say in
what goes in and what does not go into the project, someone who
controls the money. That person has to be there to resolve conflicts
arising from potential scope creep as the domain is explored in detail
and to provide provide the business's managers with feedback on
what is in the specification.
The goal of the workshop is to build a shared understanding of the
problem and the solution and to make sure that all the functional
gaps are identified early, not later in the iteration, when business
people might not be readily available. Senior business stakeholders
and domain experts are typically bottlenecks for information transfer
in software development, because they have both to do their own daily
jobs and participate in the development project. The workshop makes
effective use of their time, so that they can really focus on the issues
and transfer their knowledge to the implementation team. For a twoweek iteration, three or four hours are typically enough for a workshop. If you cannot explain what you want to build and get general
agreement on it in four hours, this probably means that you will not
be able to build it in two weeks.

Running a workshop
The workshop starts with a business person, typically the product
owner, business sponsor or an analyst, briefly describing a piece of
software that needs to be developed. He then explains how the code
should work once it is developed, providing realistic examples. Other
workshop participants ask questions, suggest additional examples

60

Keeping one step ahead

that make the functionality clearer or expose edge cases that concern
them. One of the main goals of the specification workshop is to flush
out these additional cases before the development starts, while business people are immediately available to discuss them.
Discussing examples facilitates an efficient transfer of domain
knowledge and provides a way for all the participants to ensure that
they agree on the same thing. It builds a shared understanding of the
domain. While discussing a use case, user story or whichever technique
you use to describe requirements, it is the job of the developers and
testers to raise all the tough questions and point out different
examples. The job of the customers, domain experts and business
analysts is to answer these questions and make sure that the other
participants have understood the answers correctly.
Instead of just filling in the blanks in the examples, customers and
business analysts should try to explain their decisions and make sure
that others understand them. The specification workshop is the perfect
chance to pass on domain knowledge to the implementation team
and we should use this chance to help everyone build a common
understanding. Knowing business reasons and understanding the
domain might not give developers and testers the power to make
business decisions, but it will give them enough knowledge to spot
when something is suspect and enable them to communicate much
more effectively with business people. Every once in a while, it also
leads to much simpler solutions that satisfy the customers' needs.
For the specification workshop to be really effective, people should
be free to challenge ideas and ask for clarifications until everyone
really agrees and understands the same things about the domain. The
specification workshop ends only when everyone involved agrees that
they have enough examples to move on, that they have identified all
important representative scenarios and covered all the edge cases.

Keeping one step ahead

To work more efficiently, business analysts may move one step ahead
and work with the customers to identify and specify important

61

Specification workshops

examples for the next month's work. This is absolutely fine, but it
must not be an excuse to skip the workshop at the beginning of this
next iteration. The examples that the business analysts have written
will be a good starting point for the workshop, but developers and
testers need to understand them, identify and suggest missing cases
and get a chance to discuss and critique the initial set of examples.
The workshop is especially important if there is no on-site customer
representative sitting with the implementation team, as it then gives
developers and testers regular planned access to domain experts. It
is often not possible to have domain experts constantly available, but
it should not be too hard to get them in the room with developers
every second Friday for a few hours.

Make sure that you get answers from

subject authorities
Mike Cohn wrote that the problem with business analysts acting as
customer proxies is that they often try to answer questions themselves
rather than researching the issue with users and customers ([11]
chapter 5). A similar situation can happen with a direct customer
representative if they are not an expert in a particular domain. On
one of my recent projects we had direct access to a payment manager
from the customer's organisation, and we relied on her completely
to answer all payment-related questions. But she did not really know
much about their security processes and when the time came to discuss
fraud detection, we went through the first cut of examples with her
and then asked her politely to verify the examples with the security
managers.
If you are at all unsure about the answer, just write the case down and
discuss it with business domain experts later. If a customer representative is not available, make sure to communicate the results with them
and get their opinion on all the examples as well after the workshop.
Try to elicit more interesting examples from them by asking whether
you have covered all the important cases.

62

Workshop output

Sometimes the business people will look bored or uninterested when

an example is being discussed. This is a huge warning signal and it
might mean two things. The first is that the example is not realistic
or not important for the scope of the system, in which case it should
be thrown out. The second is that you are not talking to the right
person. A call centre operator will have absolutely no opinion on
background fraud screening, and if you try to talk to them about it
they will probably just give you a blank stare. This is a signal that you
should find someone who is interested in background fraud screening
and talk to them about it, or get them involved in a different way.

Workshop output
The primary output of the workshop is a tangible set of realistic
examples that explain how the system should behave once the next
phase is complete. Key examples for every story or use case for the
next iteration should be discussed, including the main success scenario and key edge cases. They should be explained and written down
in enough detail so that all participants agree on what the story is,
what are the deliverables and how to verify that they are correct. To
keep the flow going, you might just take photos of whiteboards or
have someone write down examples into text documents as you discuss
them. It is not really important to have them in any specific format
or tool, as there will be a time to clean them up later (we discuss this
in the next chapter).
The key feature of these examples is that they should provide enough
information for developers to implement and for testers to verify the
stories planned for the iteration. The workshop should answer most
of the questions about the specification that developers or testers
would normally ask during the next few weeks of work. It is OK to
leave some edge cases for later as long as the discussion during the
workshop provides a framework to understand these cases. Make sure
to discuss success scenarios and exceptions, but don't waste time
specifying examples for all possible permutations of parameters. One
example for each important case is enough.

63

Specification workshops

To keep the workshop focused and keep the flow going, it is best to
keep the examples and discussion on a fairly high level. We want to
discuss what the software does, not how it does it. The discussion on
how software implements something should take place later and does
not necessarily need to involve project stakeholders or domain experts.
Although possible implementations sometimes need to be considered
during the workshop because they limit what can be done, as a general
rule of thumb try to avoid talking about implementation or infrastructure details. This will save time and keep the flow going.

Sharing domain knowledge

There is one additional output, which is less tangible but I consider
it even more important. The discussion about examples, edge cases
and scenarios that takes place during the workshop enables developers
and testers to learn about the domain. Not just learn, but ensure that
everyone involved agrees and understands the same thing. In a sense,
the process of getting to the list of examples that describe the desired
behaviour is more important than the examples themselves. Donald
Gause and Gerald Weinberg have a similar conclusion about their
process for exploring requirements. Paraphrasing Dwight Eisenhower's famous quote about plans and planning, they said that the
document is nothing, but documenting is everything ([3] p. xvi).
This note is important because people often focus on the tangible
outputs and forget about the benefits of the discussion. A crucial role
of the customer representative or business analyst during the workshop is to ensure that developers and testers gain a deep understanding
of the problem, as well as get answers to immediate questions. The
specification workshop is also a chance for developers and testers to
get clarification on any issues that they have with the specification.
The structure of examples provides a way to effectively communicate
the differences in assumptions and expectations.

64

Feedback exercises

Feedback exercises
Donald Gause and Gerald Weinberg suggest using ambiguity polls to
spot places where people disagree because of hidden assumptions and
help them reach an agreement ([3] Ch. 9). Their ambiguity poll idea
consists of selecting a metric that requires a solid understanding of
the domain to estimate. This can be, for example, performance, cost
or time to complete a task. The participants in the poll are asked to
estimate the metric independently and then discuss the variations in
results to grasp the reasons behind them. With a larger number of
participants, these polls often have clustered results where clusters
denote differences in understanding and variations in a cluster often
relate to observational differences. Although most specification
workshops I am involved in do not have nearly the number of participants required for numeric power laws to kick in, this idea is
applicable to smaller groups as well. In fact, it is already used in agile
planning and is known as planning poker. In planning poker, members
of the implementation team are asked to secretly select a card that
represents their estimate of how long it would take to develop a task.
All cards are then shown at the same time and the people who had
the highest and the lowest estimates explain their reasoning to the
group. This often leads to the discovery of sub-tasks or constraints
that the group was not aware of, or shortcuts that only a few people
in the group know. The process is then repeated until the estimates
converge. (Mike Cohn explains this in more detail in [12] Ch. 6)
A similar idea is described by Gary Klein in Sources of Power[4]. He
suggested using feedback exercises to improve understanding between
commanders and teams. In a feedback exercise, commanders first
give orders as normal, and then the person running the experiment
describes a potential unexpected event. The commander then writes
down how the team should react and team members also write down
what they would actually do. The notes are then compared and
mismatches and surprises are used as a basis for discussion on how
to improve communication.

65

Specification workshops

We can use these ideas to improve mutual understanding during the

specification workshop. Although the business people are the only
ones who can give authoritative answers and make decisions during
the workshop, other participants should be able to make good
educated guesses if they all share the same understanding. Taking the
core idea from ambiguity polls, we choose examples that require a
deep understanding of the domain and ask the participants to work
out how the system should behave and then explain the reasons behind
their conclusions. The differences and mismatches can be used as a
basis for further discussion and the building of a shared understanding
of the domain. If you are evaluating a numerical output with a larger
group, clusters of values might start showing up. The spread of
estimated values tells us, in a sense, how much disagreement there is.
Clusters indicate common sets of assumptions and these are also an
interesting topic for discussion.
The opportunity to run these feedback exercises offered by specification workshops is one more reason why they are much more effective
than if a single person writes examples or requirements. Gaining a
deeper insight into the business domain will enable everyone to
communicate more efficiently as the project progresses. Programmers
who know the business domain will be able to spot inconsistencies
easier and to help their colleagues resolve domain issues more efficiently. Testers learn about the domain and get involved from the
start, so when the time comes for them to verify a deliverable, they
will actually know what to expect and how to check it. This will allow
them to do their job much more efficiently. Testers can also raise
their concerns about potential problems before the development even
starts, preventing problems.

Building a domain language

One of the biggest problems for effective communication on software
projects is that programmers often come up with their own jargon.
They use class names to refer to concepts and also frequently use
technical terms such as names of design patterns and coding techniques. Business people use their own jargon to refer to the same

66

Building a domain language

concepts, often leading to a lot of misunderstanding and translation

problems. I worked on a risk management system that involved
assigning a particular set of risk assessments to an account. Business
analysts called the set of assessments a risk profile. Programmers
called it a risk template, because it was implemented as a template
that they applied to create objects. They used the name risk profile
for a particular instance of the template that was associated with an
active account. This caused a lot of weird questions and explanations
all the time.
A traditional solution for this problem is to have translators, people
who speak both jargons. A team leader or a senior programmer
eventually learns enough about the domain to be able to communicate
with business people (or the project is definitely doomed). Sometimes
a business analyst comes from a programming background so they
know that a facade is not necessarily just an exterior wall of a building,
but also the name of a popular software design pattern. These bilingual
people then become major bottlenecks in communication and their
interpretations are not always completely correct. Things often get
lost in translation.
A much better solution is to avoid the need for translation altogether
and make everyone use the same language from the start, a language
that Eric Evans calls a ubiquitous language ([13] Ch. 2). Evans suggests
intentionally selecting and evolving a language based on the domain
model, which connects team communication to the software implementation. This language should then be used consistently throughout
the project.
It is often much easier for the developers to speak the business jargon
then to teach business people about programming abstractions and
techniques, so we should start creating this language by using the
names that business people normally use for concepts. After all, the
goal of software is to solve a business problem so it makes a lot of
sense to use the current business language. This does not mean that
we are limited to existing words and phrases. The ubiquitous language
is not dead, it is a living language and evolves with the project. To use
it effectively, we will need to expand it with new concepts. The active

67

Specification workshops

use of the ubiquitous language will reveal gaps and awkward phrases,
and we then need to create new phrases in the language to cover these
cases and simplify the discussion. As we develop the model and
implement more and more functionality, new concepts will be introduced into the system and we will expand the language with names
for these concepts as well. But the important thing is to consistently
use the same language. There must be no business jargon and technical
jargon just a single project jargon, the ubiquitous language.
The ubiquitous language should be used in all the examples, diagrams,
code and speech. This involves the specification workshop, the
discussion that takes place there and any documents that come out
as a result. Examples that we use to build our specifications are a great
way to start creating the language and to challenge and evolve it in
practice. Enforcing the use of the ubiquitous language will make the
specification workshop much more effective because it helps to reduce
misunderstandings and ambiguity. At the same time, the workshop
will point to new concepts and ideas, and we will need to find names
for these new concepts. The specification workshop is a perfect place
to discuss and agree on the names so that everyone understands them.

What specification workshops are

not
A specification workshop is a very effective way to achieve a lot in a
short time, but we need to keep it focused to get the benefits. If you
notice that people have their laptops open and answer e-mail, write
code or browse the web, the workshop has failed. This might suggest
that people are simply too busy to do the workshop now, in which
case you should reschedule it. Or it can signal that you have the wrong
people in the room.

The workshop is not a meeting

It is very important to remember the goals of the workshop and focus
on them, and not let the workshop diverge into yet another meeting.
68

The workshop is not a presentation

I've organised workshops with a dozen people involved that were very
effective, but it was a real challenge to keep everyone focused on the
task in hand and not discuss other problems. This is why I think that
smaller workshops are better. Two developers, two testers and a few
business people should be enough to get the thing right. The
developers can then use the examples that come out of the workshop
to pass the knowledge on to other developers. Testers can do the same
with other testers.

The workshop is not a presentation

A workshop should be an interactive collaborative effort to produce
the complete set of relevant examples. It should not be the stage for
a single person to present a monologue on his opinions of the system.
There is nothing wrong with a single person driving the discussion,
but there has to be a discussion for the workshop to be effective.
Developers and testers should be encouraged to present their concerns
and ask questions to clarify their understanding of the domain.

The workshop is not a design session

It is OK to explore possible implementations during the specification
workshop to understand what is possible and to gain some deeper
insight into the problem, but do not get distracted and go into detailed
system design at this point. A specification workshop should be
focused on what the system should do, not how how it will be implemented. It is a very good idea to have design workshops with business
people and developers but they should be run separately and focus
on design, not scope.
Effective design workshops look further into the future and explore
different possibilities and ways to achieve the end-result when you
know what you are building. They require the participation of domain
experts and senior developers. Specification workshops help you to
decide what exactly you want to build, and they should really be
focused just on the upcoming iteration. They should include domain

69

Specification workshops

experts and senior developers as well as testers, stakeholders and the

rest of the implementation team.

How to keep the workshop focused

If you are having trouble keeping the workshop focused, think about
reducing the size of the workshop team or beginning by reaching an
agreement on what will be discussed at the workshop. Declaring one
person a facilitator of the workshop and giving this person the right
to change the topic, provoke questions and stop people if they take
over the show might be a good idea. The facilitator can then be
charged with getting everyone involved in the discussion, eliciting
questions and concerns from developers and testers and making sure
that the discussion does not go astray. I will discuss this in much more
detail in the section Choose a facilitator for the workshops on page
152.
In Sources of Power[4], Klein cites a paper by Karl Weick entitled
Managerial thought in the context of action[14], which suggests a
modified template for an effective commander's intent document:
1.
2.
3.
4.
5.

Here's what I think we face

Here's what I think we should do
Here's why
Here's what we should keep our eye on
Now, talk to me

This list of topics is a very nice reminder for the discussion that needs
to take place during an specification workshop, and it might be really
useful to you if you are trying to act as a facilitator. The steps do not
have to happen in the sequence described above and they should not
happen by a single person lecturing or commanding the rest. We want
to promote a collaborative discovery and learning effort, but we should
cover all of them in the workshop so consider the list just as a good
reminder of what to talk about.

70

Here's what I think we face

Here's what I think we face

Applied to the specification workshop, this asks for a brief introduction of the next item for discussion by the facilitator or the person
who contributed the item, explaining what the item is from his
perspective so that others in the room can understand what exactly
he is talking about.

Here's what I think we should do

Applied to the specification workshop, this asks for key examples to
be identified and discussed. This will often start by the author of the
business case explaining only the positive execution branch, sometimes
with a few common negative cases. Don't stop there. Others should
join in, identify and discuss all important examples using the techniques explained in the section Identifying important examples on
page 50.

Here's why
Applied to the specification workshop, this asks for the business
reasons behind a business case to be clearly communicated, discussed
and explained to everyone. This will provide a better framework for
understanding and may launch a series of challenges and questions
ultimately resulting in a better overall solution than the basic set of
examples would have.

Here's what we should keep our eye on

Applied to the specification workshop, this asks for potential problems
and concerns to be identified and discussed. People rarely discuss
these when planning an iteration of development. Things we should
keep an eye on includes potential problems, things that we do not
want to do (anti-goals) and possible misunderstandings, and keeping
these in mind helps a lot to focus the development on the things that
we really want to have.

71

Specification workshops

Now, talk to me
The specification workshop should not be a lecture or a talk, it should
be an open discussion. We need to make sure that everyone voices
their concerns and has their questions answered (or written down
and chased later). If some are not participating in the discussion, the
facilitator should encourage them to join in.

Dealing with open issues

Working with real-world examples and writing them down will help
flush out inconsistencies and identify gaps in requirements. This will
facilitate the discussion and help clear up a lot of edge cases. However,
some questions will inevitably be left open. The team might not have
enough information to complete the example, or they might not be
able to agree on what the system should do in a particular case. Some
proposals will require the approval of a more senior stakeholder, or
a decision from a particular business specialist not present at the
workshop. It is important to identify, discuss and write down these
incomplete questions, as well as the others. If there is a proposed
outcome that needs verification, write this down too, knowing that
it may not actually be correct. Then one of the team members (typically a business analyst if you have one) should resolve all the open
questions with a senior stakeholder or a domain specialist. Discussing
real-world examples will help business analysts communicate better
with specialists or senior stakeholders, just as it does for workshop
participants. The business specialist will not be able just to brush the
question off or give a vague answer: a real world example with precise
data requires a precise answer. Having a proposed outcome will help
the specialists review the example and decide whether it needs to be
corrected.
If a customer representative is not present during the workshop, then
the examples must be reviewed by customers after the workshop. It
is much better to have customers in the workshop, but sometimes
this is not possible. The customer should review the examples and
answer the unresolved questions. When the questions have been
72

Dealing with open issues

resolved, the examples should be updated to reflect the new insights.

A second workshop to discuss the changes is often not required, unless
the discussion with the specialist uncovers a large misunderstanding.
It is often just enough to update the expected results and have other
team members review them during the implementation.

Stuff to remember
Examples should be put to an open discussion and
reviewed by all members of the implementation team.
At the start of an iteration hold a specification workshop
to discuss examples, iron out ambiguities and functional
gaps and build a shared understanding.
Developers and testers should suggest examples of edge
cases or important issues for discussion.
Make sure that domain experts and subject authorities
answer questions. Don't make up answers yourself.
Business people should explain their answers to make
sure that others understand them correctly.
The discussion that happens during the workshop is
itself very valuable, because it teaches people about the
domain.
Organise feedback exercises to ensure that all participants share the same understanding.
Create and use an ubiquitous language consistently in
the workshop and all examples. Use the workshop to
evolve the ubiquitous language.
Keep the workshop focused, don't let it become just
another meeting.
Selecting a facilitator for the workshop can help to keep
it focused.
If a customer representative is not present at the workshop, make sure that they review and approve the
examples.

73

74

Chapter 5

Choosing acceptance
criteria
One very useful feature of realistic examples is that they are often
easily verifiable. They have precise starting values and precise expected
results. Once the software has been developed, we can actually check
whether the system pays out 14 pounds after 7672 clicks or not.
We use the examples that come out of workshops, described in the
previous chapter, to specify how a piece of software should behave
once it is implemented and confirm that it actually does what we want
it to do. Examples of system behaviour effectively become our
acceptance criteria for the current phase of development.

Acceptance tests are the specification

We try to discuss all the different interesting and important cases
during workshops. During this exploration exercise, there is often
some duplication as we uncover different cases that actually describe
the same business rules or find out that some of the parameters do
not really have an effect on the outcome. We need to clean this up,
select examples that represent all relevant differences in possibilities
and formalise them into acceptance tests. This universally agreed set
of acceptance tests becomes the specification for development. They
should be used by developers as a target for what they need to do.
They should also be be used by testers as part of their testing process.
Business analysts and customers can use them to discuss open issues
and revisit specifications when needed.
Some organisations work by requiring sign-off from key stakeholders
at various stages of their development process. In general, I consider
sign-off as a control measure that is there to ensure that someone
75

Choosing acceptance criteria

reads the specifications carefully and as a stick to beat someone with

when they want to change something later. The discussion during the
specification workshop should provide a much better review of the
specifications than any single person reading a huge document.
Introducing changes later is no big deal with agile methods, especially
if you use the examples as a live specification.
Save yourselves the trouble of maintaining and coordinating two sets
of specifications and skip writing big requirements and specifications
documents. Get clients to participate in specification workshops or
to review the examples that come out of them. If you still want or
need to get sign-off because of political or organisational reasons,
then the acceptance tests are what you should get signed off. After
each workshop, clean up the examples and print them in some formal
document layout, then get the document approved by the stakeholders
as the specification for the next iteration. If the name test is causing
problems for you in this respect, use one of the alternative names
suggested in the section Better names on page 38. It might be easier
to get a sign-off against an executable specification rather than a set
of tests.
With workshops in place, tests tend to be fairly stable during the
iterations, especially if they are focused on what the system should
do and not how it should do it. If the iterations are short enough, the
sign-off process for tests might work just fine. In practice, unless
corporate politics really prevent this, I still would not treat the tests
as something carved in stone and I would allow people to change
them during the implementation.

Choosing the right set of examples

Our goal in each specification workshop is to explore various cases
and seek to demonstrate differences in possibilities in order to facilitate discussion and learning about the domain. As suggested in the
previous chapter, throwing in numerical edge cases, very small or
large values and discussing challenges that testers or developers put
forward may lead to interesting new examples, or it may just reinforce

76

Choosing the right set of examples

an already known rule. Therefore, the workshop often produces a

number of examples that demonstrate identical rules. To make tests
easier to read, understand and verify, we should select one representative example out of each such set and add it to the set of tests. If
testers suspect that particular specific values are likely to cause bugs,
add these examples as well, even though they might demonstrate a
rule that has already been added to the tests. Having these as tests will
make sure that the bugs do not appear in the code.
Naresh Jain has adapted the SMART acronym, typically applied to
project goals,1 for acceptance tests. Good acceptance tests, according
to Jain, have these properties:2

Specific they are explicitly defined and definite

Measurable the result is observable and quantifiable
Achievable they describe a realistic scenario
Relevant they are related to the particular story
Time-bound the result can be observed almost instantly

Formalising examples into tests is the final phase in which we should

iron out any ambiguities and inconsistencies. Writing things down
in a formal way always helps to identify gaps. Ideally this should be
done at the end of the specification workshop while the workshop
participants are still in the room and ready to discuss any new issues.
Alternatively, someone may take this task offline and write the set of
tests based on examples discussed during the workshop, photos of
whiteboard drawings and things that were written down on walls or
paper. If the examples are formalised into tests after the workshop, I
strongly recommend sending the final set of tests back to all the
workshop participants and asking them to review it.

1
2

http://gojko.net/2006/10/22/magic-of-goals
http://www.slideshare.net/nashjain/acceptance-test-driven-development-350264

77

Choosing acceptance criteria

Distilling the specifications

In the section Identifying important examples on page 50 I said that
the question how would you test this piece of software is a very good
starting point for the discussion during a specification workshop.
When the result of the discussion is formalised into a set of acceptance
tests, we don't really want to record all the details of the discussion.
In order to get the most out of acceptance tests, we need to distil what
should be done from the discussion and not really focus on how it is
going to work. This is a very subtle distinction, but one that is crucial
for effective agile acceptance testing.
For example, consider this test script:
1.
2.
3.
4.
5.

A new customer registers

The customer places three books in the shopping cart
The customer goes to check-out
The customer fills in a UK delivery address
The system offers free delivery to the customer

The script describes how something is tested. But it is not really clear
what exactly is being tested here. A fairly good guess would be that
this script verifies rules for free delivery, but what are those rules? Is
free delivery offered for the first order to the new users? Or is it offered
when people buy more than two books at the same time? Maybe it is
offered to the customers from the UK? Or is it a combination of those
three things? Maybe this example does not describe free delivery rules
at all, maybe it demonstrates the general flow of activities in our online
store. Now compare the script to the following statement:
Free delivery is offered to UK customers when they place their first
order
This is the same example, restated better and focused on the really
important pieces of information. It is much shorter and easier to
understand. It does not leave so much space for misunderstanding.
If a developer was given the first test as a target for implementation,
78

Distilling the specifications

he might focus on something completely irrelevant such as the number

of books in the cart. In the case of the second test, the situation is
completely clear.
Adding more examples to the story, such as another order for the
same customer from the UK, or first orders of customers from the
US or France, will only make the first test script harder and harder
to understand. The second test stays fairly easy to comprehend:
Free delivery is offered to customers from the UK when they place
their first order
Free delivery is not offered to customers from the UK when they
place their second order
Free delivery is offered to customers from France when they place
their first order
Free delivery is not offered to customers from the US when they
place their first order
This example is a bit simplistic and the difference between these two
approaches is often much greater on real projects. Test scripts spanning a few pages might be easily summarised in a sentence or two
without losing any of the important details. Instead of having to read
and understand dozens of pages, developers and testers can focus on
a paragraph of text. That makes it much more likely that they will hit
the target when they implement or verify the functionality.
A very common symptom of problems is to specify business rules
using a workflow, such as in the case of the free delivery rules in the
previous example. Workflows in acceptance tests are so often there
by mistake that I look at any workflow tests now with serious suspicion. There are genuine workflows, of course, but much more often
a series of steps describes how something should be tested rather than
what is to be tested.
In addition to making it harder to focus on important issues, test
scripts that describe how over-constrain the implementation. By
specifying how something should be done, these tests don't allow
developers to find a better solution for the same problem. If

79

Choosing acceptance criteria

specifications only cover what should be done then developers have

more freedom to implement good solutions.
Apart from the benefit of clear and efficient communication, distilling
specifications from scripts also has significant technical benefits.
Acceptance tests written as specifications rather than scripts are much
more stable and much easier to maintain in the future. If we need to
change this rule and offer free delivery to the first three orders,
changing all the related scripts would be laborious and error-prone.
It is much easier to modify the specification when everything is
focused and listed on the same page. Script tests are more error-prone
if a step in a script changes and it was not really important for the
goal of the specification, the test will fail and the whole script will
have to be updated and reverified. If the rule was expressed directly,
those changes would not affect the test.
Rick Mugridge and Ward Cunningham wrote that many people
become disillusioned with automated testing because of the pain of
maintaining workflow tests, and that such problems typically signal
that business rules are not being expressed directly ([9] p. 133). This
is true not just for workflows, but for any test scripts that do not
specify rules directly. David Peterson was the first, as far as I know,
to point out the difference between scripts and specifications clearly,
calling for people to write specifications, not scripts.3 This is one
of the core ideas behind his acceptance testing tool Concordion.
Distilling the specification from the examples and the discussion
during the specification workshop is the key to creating acceptance
tests that are a good target for development and testing.

Specifying business workflows

Not all workflows are bad and sometimes scripts are a useful tool as
well. With examples that describe true workflows, such as page
transitions in a web application or a time-based messaging system,

http://www.concordion.org/Technique.html

80

Specifying business workflows

it makes a lot of sense to describe the example in the form of an

English language text or even a diagram.
Formalising workflow acceptance tests in the BDD format givenwhen-then (see the section Working with business workflows on page
57), is a good practice that will help you produce better tests because
it limits each test to only one setup-action-check sequence. Such tests
are easy to maintain and understand.
In the section Dealing with processing workflows on page 54 I touched
upon the subject of how to analyse and specify examples for larger
workflows, in which steps relate to individual business rules. The
main emphasis in the discussion and resulting examples should be
on the individual business rules, to avoid the combinatorial explosion
of all important cases. However, to make sure the overall workflow
is explained and understood, and that the steps are executed in the
correct order, it often makes sense to have one or two acceptance tests
for the whole larger workflow to verify key paths through the steps.
It is very important not to just take and copy acceptance tests for
individual steps and create a monster test that is unreadable and
unmaintainable. The goal of these few overall workflow tests is just
to specify how parts are connected and verify that they are connected
properly. The overall test should not be used to specify how individual
parts work in all possible cases. Individual acceptance tests or test
suites should be written to specify the functionality of individual
parts. The workflow process itself will probably be complicated, so
don't make it even harder to comprehend by complicating it more
with unnecessary details. Just take a really simple case and explain
how the system should handle it overall. The test should fit into the
given-when-then template, with all the relevant actions in the when
part. Resist the urge to verify the results of individual actions. This
functionality should already be specified and verified by more focused
business rule acceptance tests, so you don't actually need it in the
overall flow test.

81

Choosing acceptance criteria

Converting examples to tests

Formalising examples into tests is also a chance to clean them up and
focus on important aspects. Specification workshops are intended to
spark informal discussions and allow us to explore different paths
and examples. The examples that come out of these discussions often
have more details than are required for particular tests.

Each test should specify only one thing

When formalising examples into tests, be careful to identify individual
rules and specify a separate test for each of them. When several rules
are described in a single test and the first rule changes, this will affect
the results and validity of the rest of the test. Maintaining such tests
becomes a pain because we need to analyse and rewrite parts that are
not related to the particular change and then reverify the tests with
business experts. If the same example is formalised into several
smaller and focused tests, then only a single test is affected by a single
rule change. For example:
1. Mark is a VIP customer. He adds $50 worth of flowers to the
shopping cart and goes to checkout. On checkout, he is offered
free delivery since he is from the US. He can select free delivery or
keep it for the next purchase.
2. Then Mark completes the purchase, and starts again. Now he adds
a $40 flower arrangement to the shopping cart and goes to checkout, but he is not offered free delivery since he did not order at
least $50 worth of flowers. If he had saved the free delivery voucher
for the next purchase, he will still be offered free delivery. If not,
he will only be offered normal delivery.
The second step in this example continues the first one, but verifies
a different rule. In essence, completing the purchase or the total cost
in the first step are not really important for the second step. The only
important thing is whether someone has an unused free delivery offer.
So we can rewrite this as several independent verifications:

82

Converting scripts to specifications

Mark is a VIP customer. He adds a $50 flower arrangement to the

shopping cart and goes to checkout. On checkout, he is offered free
delivery since he is from the US.
Mark is a VIP customer from the US and has a previous unused
free delivery offer. He adds a $30 flower arrangement to the shopping cart and goes to checkout. On checkout, he is offered free
delivery.
Mark is a VIP customer and has no previous unused free delivery
offers. He adds a $30 flower arrangement to the shopping cart and
goes to checkout. On checkout, he is offered regular delivery
regardless of whether he is from the US or not.
Trying to specify several things at once is also a common problem
with tests described as workflows, so this is one more point to watch
out for if you are using scripts rather than specifications.

Converting scripts to specifications

You can often simplify a large number of similar scripts tests into a
single simple specification. Watch out for sets of related examples
described by similar scripts that differ only in some values in the story.
Combine all the examples into a single test that only depends on the
changing values and specify the underlying rule directly as a single
table of argument values and expected rule outcomes. Cut to the basics
and throw out all the things that do not really influence a particular
rule. Here is what the test for the three workflows in the previous
section would look like:

83

Choosing acceptance criteria

Tests should be self-explanatory

David Peterson suggests4 that acceptance tests should be selfexplanatory with a plain English description of the rule at the top
followed by the examples, that examples should only demonstrate
the rule and not include any extraneous details, and that they should
not be merged into the description. An example of a good acceptance
test is shown in Figure 5.1.

Figure 5.1. A good acceptance test

Focus the test on the rule being tested

A test should have just enough detail to provide a full context for
understanding and verification, and no more. If you are demonstrat4

in a private e-mail

84

Remember to investigate parameterisation

ing a special bonus rule that is applied to shopping carts with more
than 50 products, but the details of these products are irrelevant for
the rule, then do not list 50 products by name in the test. Write a
single step that will populate the cart with a number of random
products, in this case 50. Even in cases when you do want to list all
the products, keep the attributes actually being set in the test to a
minimum. Leave out everything that is not relevant for a particular
test. This will make tests easier to read and easier to maintain. If you
later remove an attribute from the product class, you will not have to
go through 50 examples and clean them up manually. The more
irrelevant information you strip from a test, the longer it will stay
valid because change requests and modifications of the parts that
were removed from the test will not affect it.

Remember to investigate parameterisation

One possible problem with focusing on real-world examples is that
sometimes we lose sight of the big picture. With rules such as the one
about 50 products, it is important to discuss with customers how the
system will be parameterised and how likely the value is to change in
the future. Hard-coding limits and thresholds often makes little sense,
so we should describe the rule in a more abstract way but include
realistic examples for concrete values. This will allow the system to
evolve more easily in the future.
This becomes even more important for applications that support
several customers with seemingly different business rules. I do not
like the idea of customer-specific acceptance tests, because the
customer deployment is often just a particular configuration of the
system. I've seen a case where the team activated functionality specific
to particular customers by checking configuration parameters linked
to the customer name. For example, if we had two customers called
Big Sports and Live Betting, parts of the system were activated based
on values of configuration parameters IS_BigSports and IS_LiveBetting. There are several problems with this approach. First, after a
while the second client wanted some of the same functionality as the
first client, so the team had to duplicate verifications and complicate

85

Choosing acceptance criteria

the code for these cases. Second, these particular configurations

change over time and new clients could sign up who want some
functionality from both systems, leading to IS_OurNewClient parameters that are a mix between the previous two. Instead of creating
customer-specific functionality, always think about actual parameters,
thresholds and switches and then build rules that change depending
on these switches. Big Sports and Live Betting configurations should
be sets of parameter values rather than parameters themselves. Instead
of creating customer-specific acceptance tests, work out how the rules
are parameterised and then include realistic examples for different
sets of parameter values.

Automating tests
Toyota spearheaded the success of the Japanese motorcar industry,
changing the image of Japanese cars from cheap junk to reliable highquality vehicles. One of the major forces behind this success was their
innovative production system, which included a specific attitude
towards product quality called zero quality control, described by
Shigeo Shingo[15]. Zero in this case applies not to quality, but to
quality control. In the Toyota Production System, quality was
something that was created from the start and planned in, not
controlled at the end. The point of testing in zero quality control is
to ensure that items are produced correctly to prevent defects, not
to discover them at the end. This is how we should look at the tests
in agile acceptance testing. They are there to prevent us from doing
something wrong in the first place rather than to catch mistakes at
the end.
Zero quality control introduced the concept of source tests and
mistake-proofing devices. Mistake-proofing devices are used to check
whether a part is defective right there on the production line. A key
feature of these verifications is that they are inexpensive, so that parts
can be inspected frequently. A worker can verify a part on the
manufacturing line, then it can be verified again before assembly and
again later in the process. In knowledge-based work, such as software
development, time really is money. The biggest cost on software

86

Keep tests human-readable

projects is often related to the time people spend on different tasks.

Acceptance tests are our mistake-proofing devices. In order for them
to be effective, we have to be able to execute them frequently and
quickly, without taking too much time from developers, analysts,
testers or anyone else. The best way to achieve this is to automate as
many tests as possible. Having acceptance tests in an automated form
allows us to easily check whether the code is on the right track as often
as we want.
There are lots of test automation tools available today, and your choice
of automation tools will dictate how the tests are written, what you
can automate and how to automate them. I discuss some of the more
popular tools for automating agile acceptance tests in Chapter 10. In
this chapter, I focus more on the general constraints of automated
acceptance tests. If you are browsing through this book and not
reading pages in sequence, you might also want to read an argument
against automation in the section Adopting in phases on page 151 after
reading this chapter.

Keep tests human-readable

The examples that we are turning into acceptance tests are, first and
foremost, communication devices. We use them to build a shared
understanding of the domain, but then they become a live record of
how the system should behave, and ultimately how the system does
behave once it has been implemented. When formalising examples
into tests, we must take great care to keep them in a form that is easily
understood and can be used as a basis for later discussion, especially
if we need to get them signed off by a senior stakeholder. For acceptance tests to be effective, they have to be automated but also have to
be human readable. And human in this case also includes those who
cannot decrypt the Matrix code on the fly.
If you choose to use an automation tool that makes tests understandable only to programmers, then you will have to maintain the test set
in some other shape such as a Word document. Maintaining two
separate forms of the same thing can cause consistency problems,
and violates one of the basic rules of good software development
87

Choosing acceptance criteria

don't repeat yourself. This is why I really prefer using tools that allow
us to write acceptance tests so that business users can understand
them.
Every automation tool has its own way of specifying and automating
tests, so programmers need to translate the examples chosen for tests
into code or some form of scripting. As I've mentioned before,
translation often causes problems because things get lost or become
less precise. Ideally, we want to avoid translation as much as we can.
Some automation tools allow us to write the description of a test
separately from the code automation part. When using tools like this,
it is possible to skip the translation from examples into tests altogether. For example, the test in Figure 5.2 could be automated with
FIT.

Figure 5.2. A FIT acceptance test

The automation part of this test may post actual orders or just update
a field in the database that says how many orders a customer has had
this year, depending on the design and implementation of the system.
But at this point we really don't care about how something will be

88

Automating examples with unanswered questions

implemented, just what it should do and why. By focusing on the

description of the test, not on the actual implementation or inner
workings of the system, we can zoom in on the important things. For
example, a question I'd raise is: does a threshold of twenty mean that
the twentieth order is eligible for a discount or does the rule apply
from the twenty-first order?
This separation also helps to promote specifications over scripting.
The test description can focus just on the specification and not on
how the actual test steps will be implemented. The test automation
part can hide the complexity of scripting.
The tool that you choose will drive the format of the test, so you should
ideally try to write examples in a format that the tool can directly
understand or be very close to this format, so that the chances of
introducing problems during later test automation is minimised. This
also removes the need to maintain two separate records of the same
tests.

Automating examples with unanswered

questions
Specification workshops sometimes end with unanswered questions.
Luckily, this does not prevent us from starting to automate the
examples. Lots of acceptance testing tools split tests into two parts:
the description and the automation. The description structure dictates
what the system should do, the automation part describes how the
test is executed, but individual values in the description do not
influence the automation.
For example, if the test description contains John has 50 chips in his
account, the automation code stays the same even if the user is not
called John or if he has more money at his disposal. The automation
part for the discount eligibility example in Figure 5.2 won't change
even if the customers change their minds about the case with 19
previous orders.

89

Choosing acceptance criteria

This division is done primarily to keep the test description humanreadable and shield business people from any code implementation
details. If you use a tool that separates test descriptions and test
automation, programmers can start automating tests even if some
questions are still unanswered and the tests contain some unconfirmed
results. The structure of the tests will most likely not change, even if
the values do. This often gives developers enough to start with, and
they can tweak the underlying automation a bit later to accommodate
structure changes if required.

What do we do with things that cannot

be automated?
With any set of examples, it is likely that some of them just cannot
be easily automated. As a general rule, try to automate as much as
you can, but do not be overzealous. Some scenarios with lots of
external dependencies can take so long to automate in a development
environment that it is simply more economic to give these examples
a manual run several times during development.
Performance and scalability constraints and expectations are a typical
example. A thorough stress and load test of an enterprise web system
typically requires distributed deployment on production-quality
hardware and enterprise-class storage for the database. Development
teams rarely have such hardware at their disposal during development.
Usability expectations also fall into the category of things that cannot
really be automated. Usability is a subjective quality and requires
assessment by people. Usability testing also requires observers with
relevant knowledge and experience, not readily available on most
software development teams I've worked with.
Just because something cannot be automated easily does not mean
that the ideas and practices described in the previous two chapters
do not apply. If the user interface is important for the success of the
project, then we should use the specification workshop to discuss
examples of user interfaces and make sure that they are consistent
with business rule examples. Non-functional requirements and

90

Dealing with user interfaces

specifications such as performance and scalability should also be

discussed, so that the implementation team is aware of these expectations and the members understand the constraints and reasons behind
them. Acceptance tests can be formalised for these cases as well,
although not automated, so that we know what our targets are. Then
we can verify these targets several times during the project and make
sure that the end result is fit for purpose.
This also applies to important user interface expectations and
examples. In almost all my talks about FIT and FitNesse, someone
from the audience points out that user interfaces cannot be described
with these tools and that their team is used to focusing on wireframes
and UI examples. You do not have to be constrained to a single tool.
If a user interface example is so important that it should be automated
and verified frequently, then use a different tool such as Selenium or
one of the other UI robots. Regardless of whether you automate them
or not, discuss important UI examples during the specification
workshop, in terms of both wireframes and workflows, so that
everyone understands what is expected.

Dealing with user interfaces

User interface tests can be a very effective tool if used properly, but
misusing them can cause a lot of pain and is a common source of
problems in agile acceptance testing.

Don't describe business rules through the

user interface
Customers are used to thinking about the system from the point of
view of a user of the front-end interface, so they often offer examples
in the form of user interface workflows even though they actually
want to demonstrate a business rule.
Describing business rules in the terms of the user interface is generally
not a good idea because such descriptions will be scripts, not

91

Choosing acceptance criteria

specifications. UI scripts are often too explicit and it is hard to get

through the workflow and user interface constraints to the business
rules. A lot of clicking, selecting and dragging may be required for a
simple variable assignment in domain code. Testing a business rule
properly through the user interface may require many very similar
tests, with only minor workflow differences. For example, testing the
free delivery rules in the section Distilling the specifications on page
78 through the user interface would typically require registering a
user through the web site, logging in, adding products to the shopping
cart and then going to checkout, then verifying web page contents.
With current tools, user interface tests run too slowly to be executed
often, making them a lot less effective as a guide for development.
User interfaces also tend to change a lot, at least on the projects that
I am involved in, so tests that involve a user interface are brittle and
the effort to maintain them in the long term often negates the benefits
of automation.
Many acceptance testing tools can connect directly to the business
domain layer of software and automate tests at this point, allowing
us to avoid user interface constraints and workflows and execute tests
much more quickly. Business service tests are much easier to manage
because they can be executed in a transaction and just rolled back at
the end, making it very simple to clean up after tests and ensuring
that tests are instantly repeatable. User interface tests do not give that
luxury. You can speed up user interface tests with dedicated hardware
and grid computing solutions, but they are still still a lot more
expensive (in terms of time and hardware used) than quick tests
executed directly against a domain layer on the developer machine.
For all these reasons, as a general rule of thumb, avoid describing
business rules with user interfaces.

92

Domain-specific languages for user interface tests

Domain-specific languages for user

interface tests
Mickey Phoenix presented a very interesting session called DomainSpecific Testing Languages5 at the Agile 2008 conference in Toronto
in August 2008. His basic idea was that acceptance tests should be
done through the user interface, otherwise they will require a lot of
trust from the business people, since users are not really able to see
and confirm the effects of tests below the user interface. A green bar
from a business-domain layer test tool is not the same as a web screen
popping up with the correct information. The first requires a level of
trust that the system is really doing what the test says, but the second
one makes it obvious.
Mickey Phoenix suggested creating a small domain-specific testing
language to describe user interface concepts and steps. This language
is then implemented in a general-purpose programming language
and connected to a user interface automation tool. So, for example,
a simple verification such as user is disabled might actually be
implemented as several calls and verifications across the user interface.
This implementation of a domain action would analyse web page
elements to verify the business functionality and hide all the gory
details of HTML form submission and any additional workflow
required to navigate to the appropriate page and check whether a user
is actually disabled. Having a small domain-specific language definitely helps to make user interface tests more understandable and
usable as a guide for development. This approach is similar to
xUseCase recorders, which provide record-and-replay automation
for thick client applications but allow testers or developers to name
recorded actions and use them to build larger scripts (see the section
TextTest on page 183 for more information).
This idea does have some appeal, but I believe it is still much better
to avoid user interfaces when describing business rules at the moment,
at least for the kind of projects that I am involved with. If the
5

http://www.solutionsiq.com/agile2008/agile-2008-domain.php

93

Choosing acceptance criteria

complexity of a system is in the business domain, it makes much more

sense to automate tests at the business service level. Again, don't take
these ideas as something carved in stone your projects might require
a different approach if the main complexity of the system is in the
user interface.
Although I do not agree with Mickey Phoenix's focus on user interface
tests, I am mentioning his ideas here because the concept of testing
through the user interface to make the effects obvious without
requiring any level of trust or abstraction from the customers is
certainly appealing. A major obstacle to this idea is poor tool support,
but agile acceptance testing is an emerging practice and we do not
have all the answers yet. As I explained in the introduction, poor tool
support is one of the biggest problems generally at the moment, and
I expect much better tools to appear soon. Perhaps someone will come
up with a much better web testing tool that would allow us to execute
user interface tests faster and write and maintain them more easily
and so make this idea more practical.

Specifying user interface workflows

If you want to apply the ideas of agile acceptance testing and
specification by example to complex user interfaces, make sure that
you focus the tests only on the things that really sit in the presentation
layer. Workflow and session constraints are good examples of this.
For example, a user interface test might verify that the shopping cart
summary is presented first, then the delivery options, then the order
confirmation screen. Leave the details of delivery options offered to
clients for a business rule test.
Tools such as Selenium (see the section Selenium on page 187) allow
us to automate browser actions and implement user interface tests.
As I mentioned in the section Distilling the specifications on page 78,
describing a test using low-level workflow steps and actions is not the
best idea, so you might want to abstract it a bit and make the tests
easier to understand and maintain.

94

Who should write acceptance tests?

Simon Stewart suggests using the Page Object6 pattern to make user
interface tests easier to maintain. In this approach, we describe web
pages and the logical operations they support with Java objects and
a fluent interface, hiding the complexity and details of user interface
commands in their methods. Tests are then written against the page
objects, not directly against the APIs of the test automation tool. This
approach makes it possible to write tests before the user interface is
ready, because we can work with page objects and logical methods
that will later be connected to a web site. It also means that tests are
less affected by user interface changes, since we only need to change
the page object when a particular page changes. Page objects are
essentially another version of the domain-specific testing language
idea, but in this case the domain is the workflow of a particular web
site that we are building.

Who should write acceptance tests?

In more traditional processes, a business analyst is typically responsible for writing specifications and communicating with clients, so a
business analyst is a logical candidate to identify the examples in
conjunction with the customers and write the acceptance tests based
on them. Developers have to automate the tests and connect them to
the code, so some companies charge the developers with writing tests
and then communicating them back to the analysts. A third choice
is to get testers to write acceptance tests, which is in a linguistic sense
logical.
All these approaches are, in my opinion, wrong. If developers write
tests, then they often turn out to be too technical and action-oriented.
If the business analysts or testers write them, then they often have to
be changed so that they can be automated. A better solution is to write
the tests collaboratively.
The specification workshop is a good time for developers to influence
the way that examples are written down and reorganise them straight
6

http://code.google.com/p/webdriver/wiki/PageObjects

95

Choosing acceptance criteria

away so that they can be automated and connected to the code more
easily. For example, if FIT or FitNesse is used as the automation tool,
a developer should think about fixture types that can be used to
automate test steps and may suggest rewriting some parts to make
the automation easier. Customers, business analysts and testers are
there as a safety net to prevent developers from going too far and
making the tests too technical.
If you do assign a single person to write the tests, make sure that you
know what you are doing and why you are doing it. This person is
put into a position of significant power and responsibility, as acceptance tests are the specification of the system. If you would not trust
someone to write the specifications, don't trust them with writing
acceptance tests either.
I heard about a case where the task of writing acceptance tests was
assigned to a junior external tester, just because he had some previous
exposure to FitNesse. A specification workshop was not held. A
developer with no previous experience of writing FIT fixtures was
told to automate and review the tests. These two guys just wanted to
get the job done, so they wrote the tests the best way they could,
making up the content themselves. I have a hard time imagining a
worse way to write acceptance tests. Without any domain knowledge
or understanding of the problems and requirements, they had absolutely no chance of capturing the real rules and constraints. They
made up their own theoretical examples, so the chances of identifying
any functional gaps were very thin. Introducing agile acceptance
testing in this way can only lead to disappointment, because all the
additional work will bring absolutely no benefits. Acceptance tests
written by people who do not understand the business and don't have
any influence on the scope are completely useless and your team is
actually much better off without them programmers will at least
not receive wrong specifications.
If a single person is charged with writing the tests, this person must
understand that her job is to research realistic examples and interview
domain experts on expected behaviour, not to make up theoretical
examples and definitely (and this is of crucial importance) not to

96

What acceptance tests are not

provide the answers herself. The job is probably suited best to a

business analyst, but she must communicate the tests to the rest of
the team and ask for feedback on examples and suggestions for
interesting edge cases from developers and testers. Unless such cases
are identified upfront, development will suffer when functional gaps
are identified later in the process.
Donald Gause and Gerald Weinberg suggest that a technical review
meeting is the main tool that ensures that traditional requirements
contain all and only reliable information ([3] Ch. 20). Organising a
technical review of acceptance tests is probably a good idea if a single
person writes them, but an even better idea would be to write the tests
collaboratively during a workshop. The workshop can serve as a review
meeting as well. Gause and Weinberg also suggest scrutinising the
final list of their black-box requirements tests (which relate to
acceptance tests in our case) with a team that includes members of
the implementation team and professional writers ([3] p. 258). The
review should be specifically focused on looking for over-constrained
descriptions that might impede good design and then revising them
to give more freedom for designing the solution. The professional
writers are there to improve the clarity of the specifications written
down as acceptance tests. I have never tried to do this myself, but it
might be an interesting idea to try out if you have professional writers
available.

What acceptance tests are not

A set of acceptance tests replaces the requirements and specifications
of traditional processes, but it is not used in the same way. I want to
point out some very important differences for those of you who
modify the process I describe here to suit your organisation.

They are not set in stone

Workshops do a good job of flushing out inconsistencies, but the
acceptance tests for a live software system are never complete and
final simply because the software is not complete it changes with

97

Choosing acceptance criteria

the client's business environment. Collaboration during specification

workshops leads to a much more complete set of examples, but some
functional gaps and ambiguities usually get through to the development phase. I've done no scientific research to give you exact figures,
but from my experience workshops produce significantly more
complete and accurate specifications then a single person in a traditional process. Working with realistic examples makes people think
harder, so they are less likely to change their mind.
This means that, with agile acceptance testing in place, there will be
far fewer changes than with a traditional process, but we have to
permit changes to happen to achieve complete customer satisfaction.
A large majority of changes after workshops are typically driven by
business people changing their minds (I thought about the thing that
we had on the whiteboard yesterday, and I actually think that it should
be like this... or I talked to Steve about the edge case, he does not
agree ). Tests are a live specification of the system and they can change
after workshops. Don't just take the approved set and hold on to it as
if it was set in stone.

They are not handover material

There should be no handover of specifications from business analysts
to developers. If someone needs to own the tests because of the
organisational culture, then the business people should own them
throughout the project. Developers should not be able to change tests
without telling anyone. The job of business analysts is not complete
when the tests are printed and signed off. Tests serve as a focus for
development and to facilitate future changes of software, which again
require the active involvement of business people. I cover these topics
in the next two chapters.

They are not mechanical test scripts

Tests are executable specifications, but executable here refers to a
potential to be executed, not to a primary goal of tests. One of the
most common problems with acceptance tests is to put too much

98

They are not mechanical test scripts

emphasis on execution. All members of the implementation team are

prone to this tendency. Developers sometimes go too far with making
tests easier to maintain and execute, making them very technical and
not understandable to testers or business people. Business people and
testers, on the other hand, are fond of using copy and paste operations.
Once they understand how a tool works and they have a good example
for a partial specification, they are quite capable of producing very
detailed and precise descriptions spanning hundreds of rows or
columns in a table or incredibly complex flow scripts. This is often a
result of trying to make things very precise and not thinking of
examples as something that will be read later.
It is easy to fall into this trap if any single person is charged with
writing and maintaining acceptance tests, because different roles have
different concerns about tests and benefit from them in different
ways. Many testers, especially if they come from a record-and-replay
testing background, don't really care about the inner details of a test
script, they mostly worry about the result being green when it
completes. Business people want to describe the system as precisely
as they can, using a template that developers understand. Developers
want a detailed specification, but they frown on duplication and
having to maintain hard code to integrate acceptance tests. This is
why acceptance tests should ideally be owned by the whole team.
This does not mean that complex tests that cover a full range of
different permutations are not worth writing, just that they should
not be included in the basic acceptance test suite. Nothing is
preventing you from creating an automated test that verifies hundreds
of permutations and checking the specification of this test with your
customers. Such a test can be a very useful functional regression
testing tool but it is not a good choice for an acceptance test.
The weird thing about this and the problems described in the previous
two sections is that where the problems exist, acceptance tests can
still be functionally correct and executable, we can still run them and
get a green or red result, but they are absolutely useless as a communication device. So they miss the main goal facilitating communication. Remember that acceptance tests only work when you can use

99

Choosing acceptance criteria

them for discussion. If a test page becomes too complicated, think

about describing the same test from a higher level of abstraction or
break it down into pieces that are easier to comprehend.

They are not a complete verification of the

system
I have seen people get overly enthusiastic about acceptance testing
and start to rely on them completely for testing. This does not work
and always leads to disappointment in the end. Automated acceptance
tests are a very powerful tool, but they are not a silver bullet. They do
not and should not replace exploratory tests through the user interface. Most acceptance tests execute against business domain services,
not through the user interface, so small problems in workflow and
UI layers can creep in even if all acceptance tests are passing. These
issues might be small and easy to fix, but a lot of them can render the
system unusable from the user's perspective and be very embarrassing.
Such issues can only be discovered with a human actually looking at
the user interface, not with automated acceptance tests. Automated
acceptance tests do help in the sense that they leave us with much
more time for exploratory testing, since they assure us that the
underlying business logic is working correctly.
The primary role of acceptance tests is to facilitate the transfer of
knowledge and help us build a shared understanding of what the
system should do. They are not a quality assurance tool, even though
they can help with QA.

Stuff to remember
Select a set of representative examples to be the
acceptance criteria for the next phase of the project.
Clean up and formalise these examples as acceptance
tests.
Write tests collaboratively, ideally at the end of the
workshop.

100

They are not a complete verification of the system

Write the examples in the workshop in a form close to

what your automation tool can understand.
Each test should specify a single business rule and be
focused only on this rule.
Keep tests in a form that is human-readable.
Examples should demonstrate the rule only and not
include any extraneous details
Acceptance tests should be self-explanatory with a plain
English description of the rule at the top followed by the
examples
Prefer specification over scripting. Describe what, not
how.
Acceptance tests are the specification. If you need to get
a sign-off, get a sign-off on them.
Acceptance tests guide development.
Acceptance tests are used to prevent defects, not to
discover them.
Automate acceptance tests to make them easier to verify.
Not automating everything is OK.
Using more than one tool for automation is OK.
Avoid describing business rules in terms of the user
interface.
People can add and change tests if they find functional
gaps or if a domain expert changes his mind.
Acceptance tests should not be handed over. They remain
in the joint ownership of the team.
Acceptance tests only work when we can use them for
discussion.
Acceptance tests are not a complete verification of the
system.

101

102

Chapter 6

Focusing development
With traditional abstract requirements, many details are often left to
developers to work out. Although business analysts would probably
not agree with me at this point, from the perspective of a developer
who was typically charged with the task of digesting 500-page documents and then working out what actually needed to be implemented,
I can assure you that this is true. Agile acceptance testing helps a lot
in this respect, because it gives us acceptance tests that were collaboratively produced to tell developers exactly what is needed. Once the
acceptance criteria for the next phase of the project are captured in
acceptance tests, the expected result of development is clearly specified
in a measurable and verifiable form. From this moment, the actual
development can focus on fulfilling these requirements.
A short note before we continue: although I generally try to keep this
book non-technical, this particular chapter may be a bit too technical
for business people because it deals with the actual implementation
process. Feel free to skip it if you are not a developer or skip over parts
that seem uninteresting.

Building software to satisfy

acceptance tests
Acceptance tests should be considered the specification for development. They should be an authoritative and reliable source of what the
software should do functionally, and developers can use them as targets
to implement pieces of code. Other artefacts such as functional design
documents, requirements lists and any other kind of specifications
can be produced if they help people think or communicate better, but
the functionality specified in acceptance tests is what ultimately
counts.

103

Focusing development

One of the ideas of agile acceptance testing is to flush out inconsistencies and functional gaps before the development of a feature starts.
The specification workshops facilitate this. As a result, developers
should have much less trouble completing their tasks than with more
traditional specifications and requirements documents. When they
start working on a particular service or domain object, most of the
requirements and descriptions of the expected behaviour should
already be specified in the form of realistic examples. This is the ideal
and sometimes things will need to change later and some functional
gaps will be identified during development, but this happens significantly less often than with traditional specifications. In practice,
acceptance tests provide a very good specification for all the required
functionality and a solid foundation for development. For a piece of
code to be considered complete, everything specified in relevant
acceptance tests should be implemented and confirmed by running
the tests. Mike Scott wrote1 that in his organisation acceptance tests
are used to measure the progress of development, in a metric called
running tested features. In his opinion any other measure does not
contribute to delivering value to the business.
Looking at this same idea from a different perspective, the specification is complete also in a sense that if something is not there, it should
not be implemented. The code should implement what the acceptance
tests expect no less and no more. Focusing the development just on
the things expected by acceptance tests helps a great deal to prevent
just-in-case code from leaking into the system. Ben Rady suggested
this rule of thumb for acceptance test coverage:2
And if you've got code that's not covered by acceptance tests,
you need to ask yourself this question: Can I delete this code
without affecting the functionality of the product? If so, you
should...simpler is better. If not, then you should probably write
some acceptance tests if you want to ensure that:
1. The customers are clear about what the system does.
1
2

in a private e-mail
http://tech.groups.yahoo.com/group/testdrivendevelopment/message/24125

104

Suggesting new features

2. The system continues to do it.

Developers often focus too much on the technical capabilities of the
system and put in extra features that they suspect will be useful in the
future. Mary and Tom Poppendieck argue that this might seem
harmless, but it is in fact a serious waste ([16] Ch. 1.) Code implemented just in case it is needed still has to be maintained, compiled, tested
and managed. Without clear needs and requirements from the
customers, this code may become obsolete before it is actually used,
or it may never be used. Having the code in the system leads to more
complexity and provides one more possible point of failure. The
Poppendiecks suggest a prescription to fight this complexity as: write
less code[21].

Suggesting new features

This does not mean that developers cannot think of new interesting
features. However, instead of deciding on their own that a feature is
useful and adding it just in case, developers should suggest the feature
during a specification workshop. This is the place and time to discuss
such ideas. Then the workshop participants can accept the idea and
work out a representative example of it, or reject the idea and throw
it out of the scope. This also ties in with the discussion in the section
Communicating intent on page 29, especially the anti-goals or list of
things that we do not want to do.
The list of examples that are chosen for the formal set of tests at the
end of the workshop should be complete in the sense that it should
cover all the required differences in possibilities, so if something is
not on the list we do not need it in this phase of development. As you
get more accustomed to acceptance testing, ideas about features will
be flushed out more and more during the workshop and not during
development.
Some functional gaps or ideas might be identified later in the process,
especially when you are starting with agile acceptance testing. In this
case, developers should grab a business person and discuss the idea

105

Focusing development

during the iteration. If the change is small and you have enough time
to include it in the current iteration, you can just add to the examples
and do it. If the customer representative or a business analyst is not
readily available, or the change is too big to include it in the current
iteration, it is better to leave the discussion for the next workshop. In
any case, you should try to understand together why this example was
not spotted during the workshop and try to make sure you identify
all the examples next time. This discussion can take place during the
iteration retrospective.
It is also important to remember that acceptance tests are not dead or
set in stone. They are very much alive and subject to change.
Developers may notice that they need to clean up details or require
more information to complete the task or the clients may come up
with some more requirements or change their minds. Small changes
to acceptance tests should be allowed during development, provided
that they are communicated to everyone involved. It is not uncommon
for a test to be taken out or another test to be introduced during the
implementation. This is perfectly fine as long as everyone understands
what the changes affect. If you want to introduce a large change, think
about organising a mini-workshop to discuss the change with other
team members or waiting for the workshop scheduled for the next
iteration so that you can make sure that everyone understands the
change.

Designing software based on

acceptance tests
Test-driven design is a buzzword often connected with unit testing
and lately with agile acceptance testing. With lots of different definitions of this buzzword, people sometimes take it at face value and
expect the software design to somehow magically evolve from unit
or acceptance tests. In my view, this is a wrong expectation and it will
lead to major disappointment. Good software architecture and design
do not appear out of the blue sky. Acceptance tests are specifications,
not design documents. They are a view of the system from the

106

Acceptance tests provide design hints

customers perspective and as such they should not imply software

design.
Agile acceptance testing is not a design technique, just as it is not a
development methodology. You can theoretically use any software
design practice with acceptance tests, although in practice it only
works with techniques that support incremental and iterative design.
The way that workshops build up specifications in small stages makes
it very hard to do a precise global system design at the start of a
project. Building specifications with acceptance tests phase by phase
allows us to incorporate knowledge from previous phases of a project,
which leads to the improvement and refinement of the system design
and architecture.
From my experience, the best results are achieved when there is an
overall architectural vision formed at the beginning of the project
and then the design is built up incrementally in accordance with the
vision. Note that the word vision does not imply detailed specifications or something set in stone, but an idea that is simple and clear
enough so that people can relate to it and use it as a guideline. Software
should be designed just-in-time as new components in the system
and simplified and refined iteratively to be consistent with the overall
vision.

Acceptance tests provide design hints

Having said all that, acceptance tests can provide very useful hints
about the interfaces between business services and domain objects.
This is where the idea of test-driven design comes from.
For example, an acceptance test that specifies required payment details
can hint at the structure of our domain object for a credit card. From
the test in Figure 6.1, we can suppose that our card will need a type,
card number, cvv code, issue number, and start and expiration dates.
The test also suggests that the domain object or a related business
service will have a method to check card validity.

107

Focusing development

Figure 6.1. A test suggests domain object properties

Similarly, a workflow test could suggest business service methods and

their arguments:

Given that Mark places 50 chips on number 12

and Tom places 10 chips on number 13
When the roulette wheel spins and stops on Odd
Then Mark loses
and Tom wins 20 chips

This workflow test hints that the service would most likely have a
method to place chips on the table, with a player, amount and selected
field as arguments. It also hints that there should be a method to spin
the roulette and a method to check how much a player won.
Just make sure not to take these hints as design requirements. For
example, it would be false to assume from the fact that the acceptance
test specifies the roulette table field Odd that the Spin() method for
the roulette business service takes the result field as an argument.
Code using the roulette table should not be able to specify the outcome
of the game, but we need exactly this in a test. The link between the
test specification and the business code for this step would have to be
implemented using a test system instead of a real random number
generator, for example, but not through the business service.
Tests should be focused on the behaviour under test and they often
do not provide the full precise context. For example, if a special bonus
should be applied to shopping carts with more than 50 products, the
108

Design does not have to be feature-driven

acceptance test for this should specify a case where the cart has 50
products (as described in the section Focus the test on the rule being
tested on page 84). It would be wrong to assume that this implies a
business method which randomly generates a number of products to
fulfil the test workflow. On the other hand, this test hints at the need
to count the products in the cart, which should most likely be implemented as a domain method.

Design does not have to be feature-driven

If these ideas sound similar to feature-driven design, don't take them
as implying that you have to adopt feature-driven design in order to
use acceptance testing. I want to stress again that agile acceptance
testing has nothing to say about the way you design the software.
Other design techniques are all viable options as long as they can
work in an iterative fashion. Domain (model)-driven design in
particular is a very good fit for iterative processes. With domaindriven design, you might want to explore several examples that will
be developed in future at a less detailed level, in order to achieve
deeper insight into the domain. Agile acceptance testing does not
stop you from doing this. It does not help in any particular way, but
it does not prevent you either.
You can look at the stories for the whole delivery milestone and discuss
some of them with the business experts in order to get enough
information to build and verify your model design. These examples
should not be put into acceptance tests at the start they can be
included later when the time actually comes to develop them. You
might want to capture them in order to kick-start a later specification
workshop, but you do not have to do this.

Evolving the design

One more important idea to mention about software design is the
practice of using tests as a sort of scaffolding to hold the software
together during architectural clean-up or redesign. Refactoring is a
term coined by Martin Fowler to describe improving the design of

109

Focusing development

code without modifying the functionality. This practice depends on

automated tests to notify us if we change the functionality during
redesign and clean-up, so that we can quickly spot and correct any
problems introduced. The idea is often summarised as red-greenrefactor. The phrase describes a typical development cycle. We should
first write tests to specify what we are going to develop. These tests
will initially fail (red) because there is no code to satisfy them. Then
we write the code to make the tests pass (green). After this, we clean
up the design and improve the fit between the new code and the rest
of the system (refactor). Running the tests again allows us to ensure
that we did not break anything during the clean-up.
The red-green-refactor idea is often mentioned in the context of
unit tests, but it is actually better suited to acceptance tests. Acceptance
tests describe the functionality from the perspective of customers,
without going into technical implementation details. This customer
view of the functionality is really what must not change before and
after refactoring. Unit tests specify and later reflect a particular
technical design. When the design improves, these unit tests are likely
to break. Acceptance tests, on the other hand, should keep running
correctly regardless of the changes in the technical design.
Developers should be able to adjust and refine the technical design
as long as the functionality that customers observe stays the same.
We develop the code to satisfy the acceptance tests, then, once the
tests pass, revisit the design and make the new code fit better with the
rest of the system. Rerunning the tests ensures that the functionality
is still there after the code is cleaned up. This is how the software
design can evolve supported by tests.

Developing the glue between code

and tests
With automation tools that allow us to split the description of the
test from the automation part (see the section Keep tests humanreadable on page 87), each specification workshop ends when the

110

Use the project language consistently

test descriptions are written down. Developers then implement the

domain code and the test automation part in parallel, with glue code
to connect them. This glue code allows us to effectively separate the
specification from implementation details. Ideally, all scripting details
should be encapsulated in the fixtures, leaving the acceptance tests
focused purely on the specification (see the section Distilling the
specifications on page 78).
Because the glue code is a link between domain code and acceptance
tests that business people can understand, it often ends up being very
ugly from a coding point of view. This sometimes bothers developers
because it is contrary to what they are normally expected to do.
Changing the glue code often requires changing the test descriptions
as well. If you decide to change test descriptions, make sure that the
new descriptions are still easily understandable to business analysts
and customers.
Most tools will allow you to extract common parts of acceptance tests
into a set-up or a tear-down section, so that you can reuse them across
several tests. Re-organising tests like this to improve reuse is a good
thing, because it makes them easier to maintain. However, make sure
that you don't go too far. Resist the urge to fix the glue code and
refactor tests so that they are aesthetically nice from the coding point
of view but require big changes in test descriptions.
I have no perfect solution for this problem. I would only say that while
the glue code might be ugly and a bit of a pain to maintain, it provides
improved direction during development as part of the agile acceptance
testing practice. The trade-off delivers.

Use the project language consistently

It is very important to keep in mind during development that the
language used in acceptance tests should be reflected in the code as
well. In the section Building a domain language on page 66 I explain
how the specification workshop helps us establish a common language
for the domain. In development, we want to use this same language

111

Focusing development

for the concepts in our code and avoid creating a technical jargon for
the project.
Acceptance tests help promote the ubiquitous language concept of
domain-driven design, because they give developers an obvious choice
of names for domain elements described in the tests. Consistent use
of the domain language in examples and tests will lead to a consistent
use of the language in code as well.
Concepts in the code should be given the same names as the respective
concepts in test descriptions and in other documentation. During the
implementation of the tests, you may notice some inconsistencies in
naming. This is perfectly normal especially in the early stages of the
project, since the language is still being formed and we do not use
anything to enforce formal naming conventions during specification
workshops. In two different tests we may refer to the customer's
address as mailing address and home address. When you start
implementing those two tests, you will notice the difference because
code compilation enforces formal naming. When things like this are
found, developers should communicate with the business people and
decide which name to select, and then consistently use this name.
This also means adjusting one of the test descriptions to make it
consistent with the code.
If you introduce a new domain concept during implementation
because it is required to complete an example, but the concept does
not appear in the tests, my suggestion is to contact the business people
again and ask them to suggest a name for the concept. Don't just give
a class the first name that comes to your mind. If the concept is a firstorder domain member, the business people may well already have a
name for it. Let's use names for our classes and objects that the business people can understand too. This will make it much easier to
discuss the concepts.

112

Unit tests vs acceptance tests

Unit tests vs acceptance tests

Due to a combination of linguistic confusion and years of training
to avoid duplication in code, developers often raise the question of
whether and how they should choose between unit tests and acceptance tests. The linguistic confusion exists because of the word test.
If you look at acceptance tests as executable specifications, then the
question becomes much less confusing. Should we write executable
specifications or unit tests? Why choose between them we need
both.
Unit tests and acceptance tests focus on completely different things.
Unit tests act as a micro-target for code units and they check whether
the code is correct from a technical perspective. Unit tests should
examine edge cases such as empty strings, incorrect formats and
various combinations of input arguments that the programmer is
concerned about. Acceptance tests act as a macro-target for development of whole code modules and they verify that the product is correct
from a business perspective. They examine realistic business cases
that the customers and business people are concerned about. Andy
Dassing wrote that the unit tests will insure the code is built right,
and that acceptance tests insure the right code is built.3
When acceptance tests are used in the project, large parts of the
production code are covered by these tests. Some people tend to skip
unit tests because acceptance tests already check the functionality.
Although this practice does save some time, it may make it harder to
pinpoint problems later. Failed acceptance tests signal that there is a
problem, but do not locate the source as accurately as unit tests do.
Acceptance tests also rarely check purely technical edge cases, so unit
tests should be written to cover these issues at least. Infrastructure
parts of the code, not specifically related to any user story, are also
not properly covered by acceptance tests.

http://tech.groups.yahoo.com/group/fitnesse/message/10115

113

Focusing development

For example, an acceptance test for a search system might specify that
a user can enter multiple search phrases separated by commas in the
same line. Somewhere under the hood, this big search criterion needs
to be split into multiple criteria. Splitting the search string should be
a responsibility of a distinct code unit which is purely technical. The
overall acceptance test is not a particularly good target for development of this code unit. It will give us a green or red result, but it will
be difficult to know whether the problem is in the code that splits the
string into pieces or the code that executes the search. It will also not
give us quick feedback while we develop the functionality, as the search
needs to be developed as well in order for the acceptance test to run.
In cases like these, I prefer to write a few unit tests for the required
string manipulation functionality. These tests will allow me to develop
and test the string manipulation code unit in isolation and ensure
that the unit is correct before I include it in a wider context. Dividing
the code up like this also allows us to split the work. One developer
can work on string manipulation, another can work on executing the
search. The overall acceptance test verifies that both code units
cooperate correctly to give the correct business result at the end.
Unit testing tools are often used to perform integration tests or
exhaustive system testing. With acceptance testing in place, such tests
can be easily dropped. Having acceptance tests in the project in
addition to unit tests can help focus the unit tests on what they really
should be checking code units from the programmers' perspective.
Unit tests should not verify large workflows, business rules, connect
to a database or require any form of external system set-up or
configuration. We can move such verifications from a unit test suite
to the acceptance test suite and keep unit tests nice and quick.
Some duplication between unit tests and acceptance tests is not really
a problem because it does not introduce a large management overhead,
and you should not drop legitimate unit or acceptance tests because
the same thing is already verified in the other groups.

114

To mock or not to mock

To mock or not to mock

In order to isolate unit tests from the rest of the system, it is a common
practice to use mock or fake stub objects to simplify dependencies.
Programmers sometimes take this practice and extend it to acceptance
tests. In my opinion, this is wrong.
I like to execute acceptance tests in an environment as close to the
real thing as possible and avoid any mocking. Using mock objects in
unit tests is OK because it promotes isolation of code units. Using
mock objects in acceptance tests can hide serious integration problems. When an acceptance test passes, the development should really
be done. If mock objects are extensively used for acceptance tests, then
you will quickly lose track of what has really been done and what still
needs work.
This is why acceptance tests must connect to a real database, use real
services if possible and in general provide us with as much assurance
as possible that the thing really does what the customers want. Because
of this, acceptance tests often run much slower than unit tests.
Mocking external systems that have no real test APIs is unavoidable,
such as in the example of a random number generator described in
the section Acceptance tests provide design hints on page 107. Mock
objects are sometimes the best solution to verify time-dependent
processes or isolate very slow systems. Although they are sometimes
unavoidable, I really think that we should keep mock object usage to
a minimum with acceptance tests.

Running acceptance tests

A major difference between unit and acceptance tests is also that for
most of the time during development unit tests pass and acceptance
tests fail. When all acceptance tests pass, the development is done
because this means the requested functionality is in place.
Developers should not necessarily run all acceptance tests before
committing every change to a version control system, but they should

115

Focusing development

run all unit tests as we want to catch technical bugs before the code
goes into the central repository. For unit tests to be effective, they
have to run quickly and execute from the integrated development
environment. For acceptance tests, it is much more important that
they are easily understandable by business people and that they can
verify the system in a state as close as possible to the production
system.
Acceptance tests tend to be slower than unit tests, because they connect
to real databases and external services. Depending on external
dependencies, acceptance tests might not really be executable at all
on developer machines. Whereas this would be a huge problem with
unit tests, it is relatively fine with acceptance tests.
Even if your acceptance tests can run on developer machines, do not
include them in the basic build. They will just slow down the turnaround time for implementing small changes. Unit tests should
execute on every change to verify that bugs are not being introduced.
This basic test run should not last more than a few seconds, as it has
to be done frequently. Developers should run acceptance tests to verify
that they have finished with a larger piece of code from a business
perspective.
Automated acceptance tests are easy to run, although they typically
do not execute quickly (or at least not as quickly as unit tests).
Developers should take advantage of this and run them periodically
to verify that they are on the right track. A continuous integration
system should also run acceptance tests for previous phases of the
project overnight as a regression test to verify the system functionality.
If you do this, it is a good idea to publish the results somewhere where
the business people, project managers and customers can see them.
This will allow people to see what progress is being made.

116

Running manual tests

Running manual tests

With acceptance tests that cannot be easily automated, it makes no
sense to run them throughout the project, but it also does not help
to leave them to the end of it.
Instead of running performance and scalability tests regularly and
frequently, we can schedule such tests on rented hardware or in a
dedicated test centre several times during the project to ensure that
we are on the right track. I was recently involved in the development
of a high-throughput transaction processing system. We ran a stress
test halfway through the project, once the overall architecture started
to take shape and there was enough functionality for the test to be
relevant. The results were quite disappointing so we asked for outside
help and got a consultant to investigate and help us configure the
infrastructure properly. Once this obstacle has been removed and we
got a satisfactory throughput, we continued developing and then
executed another stress test a few months later to reverify and readjust the system according to performance expectations. Although
early optimisation is a generally bad practice, I do not like to leave
the performance testing to the end. Identifying bottlenecks and realigning the system a few times during development is usually much
better, because we learn from our mistakes and incorporate the
knowledge into further development.
The same applies to other types of tests that cannot easily be automated or cannot run on development hardware, such as security
probing and usability testing. Instead of leaving these to the end, it
is much better to examine the system several times during the project
and incorporate the knowledge into further development to promote
continuous improvement.

Stuff to remember
In practice, acceptance tests provide a very good
specification for required functionality.

117

Focusing development

A piece of code is complete when all the related

acceptance tests run correctly.
Agile acceptance testing does not require you to use any
particular design methodology.
Acceptance tests might suggest what is required, but
they do not imply design.
Resist the urge to clean up test automation code while
making test descriptions unreadable.
Use the same names for concepts in the code as the
names used in acceptance tests.
Unit tests verify that the code is right, and acceptance
tests verify that you have the right code. You need both!
Acceptance tests should run in an environment as close
to the production environment as possible.
Make sure to periodically run tests that cannot be easily
automated or run on development hardware.

118

Chapter 7

Facilitating change
Once the entire set of acceptance tests for a project phase is green,
the development work is almost done. At this point, developers might
improve the design of the system with some new insights acquired
during the current iteration and re-execute the tests to confirm that
the system is still working as it should be. After this, we can continue
with the next phase of the project. The role of the set of acceptance
tests for the previous phase now changes from that of a target for
development to that of a utility that facilitates future change. It
becomes a live record of what the system does and a set of regression
tests that can be executed to verify that the system still does what it
was supposed to do.

A live specification
The executable code is often the only thing that we can really trust in
software systems. Other artefacts such as specifications, requirements,
even API documentation, quickly get out-of-date and cannot be
trusted completely. On the other hand, executable code is unusable
as the basis for a discussion with business people. Source code is no
better, since business people cannot read it. Even if they could read
it, source code has too many details to facilitate an efficient discussion
about business rules and features. It is too low-level and it prevents
us from seeing the wood for the trees.
An automated acceptance test suite, connected to the executable code,
can easily be run to check whether it still reflects reality. We can have
the same level of confidence in the automated acceptance tests as we
do in the code. This allows us to use the acceptance tests written for
previous phases of the project as a good basis for discussion. The
representative set of examples, formalised into acceptance tests and
connected to the code by test automation, plays the role of a live

119

Facilitating change

specification of the system the authoritative source of what the

system does and how it behaves.
This kind of live specification is easily understandable by business
people and helps everyone understand the business domain. Rick
Mugridge gave a very interesting presentation called Doubling the
value of automated tests1 at the Google Test Automation conference
in London, on September 7th 2006. In this presentation, Mugridge
talked about how acceptance tests helped people understand the
business domain. Business analysts worked with testers on writing
tests, and started noticing inconsistencies and unclear definitions
when they wrote test examples. Tests not only helped guide the
development, but also allowed people to clarify the business domain
and understand the business rules better. When acceptance examples
evolve during the project, duplication is removed, workflow rules are
flushed out, and acceptance tests become focused on the business
domain and express true business rules. Mugridge says that acceptance tests became a significant and valuable business resource,
helping to educate employees on the business.
Having a live and relevant specification of the system helps new team
members to get started and understand the project. It also protects
us from the bus effect, whereby we lose important knowledge if the
only person who knows what the system does gets hit by a bus (or in
a more realistic scenario leaves the team for a different job).

Keeping it live
During development, we expect acceptance tests for the current iteration to fail most of the time, since the functionality is not yet there.
But once the functionality is implemented, previous tests should
always pass. After a phase of development is done, it is of crucial
importance to promote related acceptance tests into the suite of
regression tests. These tests can be used to verify that the system keeps
doing what it is supposed to do.
1

See http://video.google.co.uk/videoplay?docid=-7227306990557696708 and

http://www.rimuresearch.com/RickMugridgeGoogleConference.pdf

120

Keeping it live

Running acceptance tests for the current development phase overnight

and publishing the results is a nice way of informing everyone of
progress, but it's not really mandatory. However, running the
regression test suite for previous development phases periodically is
a necessity. Tests are automated, they run overnight, so there is
virtually zero cost in executing all of them. (Except in a very rare case
when the whole suite cannot be executed in one night, so you might
need to divide it into batches or prioritise tests.)
If one of the tests fails, it means the system stopped doing something
that it is supposed to do and this should raise an alarm. A failing test
that was previously running is a clear signal that the code is no longer
synchronised with the specification, so one or the other needs to be
modified. This may be due to a bug being introduced with new code,
or because of a change that conflicts with some of the previously
implemented business rules. Unless it is completely obvious that the
problem is due to a bug (in which case the code should be fixed
straight away), customers have to decide whether the example is
obsolete or still valid. If the team has a dedicated business analyst,
then it is her job to explore this issue and discuss it with customers
and stakeholders.
A very important thing to remember is that regression tests in doubt
should never be disabled. Developers sometimes disable a test to make
the alert temporarily go away, but they rarely remember to go back
and re-enable it. Tests should be disabled or dropped from the
regression test suite only if the functionality is obsolete. Do not ignore
the error signals or disable tests to make the problem go away. Either
fix it or get confirmation that the functionality specified by the test
is no longer valid. I do not advise waiting for the next specification
workshop to sort out these problems deal with them as soon as you
can.
A continuous integration system should run the tests and tell someone
if any of the tests fail this will protect against problems creeping
into the system. Automated alerts keep the problems small, since we
can catch issues as they appear and solve them, and not allow them

121

Facilitating change

to accumulate or grow. Otherwise, you will soon be affected by the

broken window syndrome.
Broken window
The broken window syndrome was first described in an article
on police and neighbourhood safety, written by James Q. Wilson
and George Kelling and published in 1982 in Atlantic Monthly.2
Wilson and Kelling argue that disorder in a community, if left
uncorrected, undermines efforts to maintain the neighbourhood. The authors suggest that a single sign of carelessness can
cause huge problems later, if not taken care of. One unrepaired
window is a signal that no one cares, so breaking more windows
costs nothing is one of the key ideas in the article. Willson and
Keeling give an example of litter accumulating on a pavement
and not being cleaned up. Soon, more litter accumulates.
Eventually, people even start leaving bags of trash from takeaway restaurants there or breaking into cars.
This article and the broken window syndrome were introduced
to the software development community by Andrew Hunt and
David Thomas in The Pragmatic Programmer[18]. They
compared broken windows and accumulating problems to issues
in the code. If a piece of code is clean and properly covered with
tests, programmers who modify it will most likely try to keep
it clean. If it already has a ton of problems, they simply will not
care. If problems are left to pile up, people stop caring about
the design and just hack the code to get their job done
breaking more windows costs nothing.
Making sure that previous acceptance tests are always valid is also
important from the aspect of keeping the specification relevant.
Because all tests are passing correctly, we know that the specification
is still live and we can trust it in discussions. The periodic execution

See [17] and http://www.theatlantic.com/doc/prem/198203/broken-windows

122

Introducing changes

of tests is like a beep on the heart monitor during medical operations,

telling us that the patient is still alive.
Because running tests periodically is so important, it should not be
left to anyone to choose whether they are executed or not and we
should definitely not rely on any single person pushing the button to
do it. Continuous integration servers, such as Cruise Control by
ThoughtWorks and TeamCity by JetBrains, automatically compile a
project from source code and execute tests to verify that it is correct.
When a problem is found, they can alert developers and testers by email or signal problems using a Web console. We should use such a
system to alert us about any inconsistencies between the acceptance
tests that were previously running correctly and our code. When you
put acceptance tests under the control of a continuous integration
server, remember that a failure in the regression test suite should
break the build, unlike an error in the acceptance test suite for the
current phase, which is not really expected to pass until the end of
the phase.
Acceptance tests are a good start for the regression suite, but don't
take this as a suggestion to stop there. Testers should use other tools
and methods for supplemental regression testing as they see fit. I want
to make this explicit so that there is no misunderstanding. Agile
acceptance testing just gives you a good starting point, because it
ensures that the business rules are covered. (Comprehensive regression
testing is outside the scope of this book.)

Introducing changes
With a live, relevant specification in the form of acceptance tests that
business people can understand, we can introduce changes to the
system much more easily. The live specification facilitates discussion
about change requests, and we can use existing tests to analyse how
the requested modifications would affect system behaviour. Existing
examples provide a good framework in which to discuss the changes
and make sure that we do not forget important edge cases. When we
start changing the existing tests and adding new information to them,

123

Facilitating change

having things written down on paper, or in an electronic form, will

again help us identify new potentially important cases or functionality
gaps (as explained in the section Identifying important examples on
page 50). These changed tests are then moved from the regression
test suite back to the active set of acceptance tests for the current phase
of development.

Solving problems
With domain knowledge and understanding shared among team
members and a comprehensive acceptance test suite based on realistic
examples, the quality of the resulting software is much higher than
with more traditional specifications and requirements. This does not
mean that you will never get any bugs. Nobody is perfect and unfortunately bugs will still happen.
When a bug is found, the key questions to answer are: why was it not
caught by the tests in the first place? and which set of tests should
have caught it? This will provide insights for future specification
workshops. Mike Scott recommends3 that we should treat defects as
evidence of missing tests. If an existing feature is incorrect, then the
case was overlooked by the team so we need to add a new test and
ensure that similar cases are covered next time. Another possibility
is that the case was specified incorrectly, which means that real domain
experts did not participate in the workshop and did not review the
tests afterwards. To solve similar problems in the future, we need to
identify people who should participate in the workshops or review
tests for each part of the system. If the system is not doing something
that it should be, then a new feature should be put into the development plan and specified when the time comes for implementation.
There is also the possibility that the bug is not a domain problem at
all and should have been uncovered by unit tests or integration tests,
not by acceptance tests.

in a private e-mail

124

Keeping the software flexible

Unless the bug is a complete showstopper and has to be hot-fixed on

production as soon as possible, I advise modifying the relevant test
or writing a new test to demonstrate the bug first. Then the developer
can have a clear target for fixing the bug and run all the tests after the
fix to make sure that nothing else was broken in the process. With
showstoppers that get fixed, deployed and verified in production, this
discussion should happen after the resolution when things calm down.
It is important to change the acceptance tests in the light of this new
knowledge to prevent the bug from resurfacing in the future!
Once a bug verification test is in the regression test suite, it will be
executed every time with the other acceptance tests, making sure that
a fixed bug stays fixed.

Keeping the software flexible

At the beginning of development, changes to software are quick and
simple. As the code base grows, it becomes harder to modify. A simple
change in one area often causes problems in a seemingly unrelated
part of the code. Without tests, it soon becomes too hard and
expensive to change anything. With tests to signal problems, we can
be more confident in the code and change it freely.
Because we are gradually building the regression test suite as we are
building the functionality, we always have scaffolding that will hold
the code together during changes. This allows us to improve the design
more easily and to simplify the code, making sure that new changes
fit into the system nicely and that we have incorporated new knowledge correctly. Periodical housekeeping keeps the software flexible
by making it easier to maintain and extend later. Running the tests
after each change and making sure that they all execute correctly tells
us that we have not broken anything and that the system still does
what it was supposed to do.

125

Facilitating change

Keeping the specification flexible

Although tests might start focused and simple, they usually degrade
over time. People will often try to minimise changes to tests when
code gets changed or when business rules change, so it is likely that
additions and changes will only be made with time. Surplus items
that do not really affect the outcome will often not be deleted.
Although new insights and breakthroughs in the design of the domain
or code changes might uncover ways to write acceptance tests simpler
than they were done originally, tests are often kept in the original
form if they still work. These tendencies mean that tests will gradually
lose their usefulness as a true, live specification, if some effort is not
put in to improve them.
Acceptance tests should facilitate change, not impede it. If you find
yourself in the position of resisting changing the system because it
will take a long time to fix all the tests, then the tests have to be
simplified and made easier to change too. This is why, like code,
acceptance tests need periodic housekeeping and clean-up so that
they remain flexible.

Common symptoms of problems

Here are some things to watch out. Try to avoid them when changing
tests, or clean up tests if you notice any of these problems:

Long tests, especially those that check several rules

Parameters of calculation tests that always have the same value
Similar tests with minor differences
Tests that reflect the way code was written or tests that mimic code
Tests that fail intermittently even though you haven't changed the
code
Parts of tests or even complete tests used as set-up for other tests
Interdependent tests.

126

Common symptoms of problems

Long tests, especially those that check several

rules
If a test is unusually long, it is often a signal that the test contains
much more detail than what is really needed to test a specific business
rule or feature. As the system evolves, a long test will be affected by
code changes much more frequently than required, since changes in
any of the details featured in the test will require you to update it.
Another possibility is that the test verifies several constraints or
business rules in a single workflow. In this second case, the test will
be affected by changes in any of the business rules and these changes
will have a domino effect. If the third step of the test has to be changed,
the fourth and fifth will probably have to be updated as well. If the
test had been focused on a single business rule, then changes to related
rules would not have affected it.
A similar problem can occur with a test that describes a state machine
or a calculation with a large number of parameters. This often ends
up as a table with a large number of columns. Most likely some of
these parameters do not really affect the calculation or the state. They
might have been important earlier but became obsolete, or maybe
they were just copied from a similar test, but never really affected the
behaviour specified by the test.
Acceptance tests can become too long because a number of small
changes or related verifications are bolted onto existing tests, as it is
easier to do this than create a completely new test. You might make
several changes to the test as the code is refined before you spot that
it actually describes two or three different things.
In any of these cases, you should look for ways to make the test shorter
and simpler. If it specifies several business rules, split it into smaller
and more focused tests (see the section Each test should specify only
one thing on page 82). If it contains extra parameters that are not
important for this particular test, delete them. If it is too technical or
detailed, try to specify the test from a higher level of abstraction. In
particular, move any implementation or scripting details from the
127

Facilitating change

actual acceptance test description to the automation code. An

acceptance test should ideally specify what the system does, not how
it is implemented (see the section Distilling the specifications on page
78).

Parameters of calculation tests that always have

the same value
When a parameter in a description of a state machine or a calculation
rule always has the same value (or keeps the same value in a majority
of cases), this is typically a signal that the parameter is not really
important for the feature described in the test. If the value is there
just because the underlying business method under test needs that
argument, then the test automation code can generate a random value
or pass a hard-coded value, allowing us to drop the parameter from
the test description. This can also happen if the argument was once
relevant for the calculation but the rules changed, making it obsolete.
Surplus parameters and repetitive code make it harder to maintain
a test.
If the value is really important for the calculation, it can be set globally.
For example, instead of a table that verifies tax calculation on products
with 50 rows with only two types of products mentioned in these 50
rows, we can split the test into two tests of 25 rows, each with a globally
specified product type, and remove the product type column from
the tables.

Similar tests with minor differences

Lots of similar or related tests with only minor differences are typically
a symptom of a business rule explained through scripts rather than
directly. Tests such as these demonstrate how the underlying business
logic works, rather than what it is supposed to do. Having different
cases for a single business rule broken down into several scripts with
just minor differences will make it harder to see the big picture and
identify additional important edge cases. This set of tests will be a
problem for maintenance because the tests most likely include some

128

Common symptoms of problems

steps that are not necessarily relevant for the thing being described.
Changes in these seemingly unrelated steps will affect all the tests as
well.
This might happen because we started with a single story and then
added similar stories as change requests came in. Alternatively, the
stories might have initially started off as different but become similar
after code refactoring or test clean-ups.
In general, a good solution for this is to convert all the script-style
tests to a single focused specification test. This test should only contain
the relevant arguments and specify what the underlying functionality
does rather than explain how it works. You can see an example in the
section Distilling the specifications on page 78.

Tests that reflect the way code was written or tests

that mimic code
Acceptance tests should be focused on the customers' view of the
system. When tests are too action-oriented or mimic code, they get
cluttered with too much detail and make it hard to see the big picture.
In addition, they will need to change when the implementation
changes even if the underlying business functionality does not.
Tests like these often get written by programmers. Sometimes they
are written after the code to increase the code test coverage. This is
very dangerous because it is a sign that programmers are changing
the specification of the system to match their view, which might not
necessarily be the same as the view of the business people. In addition,
such tests are often biased towards the implementation. If there is a
problem in the implementation, the test hides it and makes it seem
as if the system is correct. Any tests written just to increase the code
test coverage should be reviewed by domain experts. Tests that are
too technical and action-oriented should be rewritten to use a greater
level of abstraction.
This does not mean that tests with a low level of detail focused on
implementation should not be written. But these tests should be in

129

Facilitating change

the set of (programmer) unit tests and not in (business) acceptance

tests. As a general rule of thumb, developers should not be writing
acceptance tests themselves. Acceptance tests have to be specified
collaboratively with business people and testers.

Tests that fail intermittently even though you

haven't changed the code
Acceptance tests (and unit tests for that matter) have to be reliable.
If they pass, we must know that the system works. Once they get
promoted to the regression test suite and fail, there had better be an
error or we are going to waste a lot of time investigating for nothing.
Once the problem has been found and fixed, we need to be able to
trust that a green test is a sign that the job is done. Tests that fail
intermittently without underlying code changes are misleading and
unreliable we are better off without them.
Tests that sometimes pass and sometimes fail are a huge warning
signal. How do we know that the same thing is not going to happen
in production? Many causes of unreliability in tests, such as timing
issues, dependencies on database contents and reliance on random
values, can be fixed. If the issue is caused by an asynchronous process,
we can focus on testing the synchronous parts in isolation for
acceptance and then just have a simple overall test to check
connectivity. If tests depend on database contents or constraints, they
should be wrapped in transactions which are rolled back after the
test.4 Tests that depend on random values might need to be rewritten,
or the random value generator might need to be abstracted to make
tests easier to write (this is a rare exception where a test mock is useful
in acceptance testing see the section To mock or not to mock on page
115).
Regardless of what the cause is, tests like these have to be investigated
and fixed, otherwise they will undermine the reliability of regression
tests in the future. As a consequence, these tests also undermine
confidence in the code under test.
4

http://gojko.net/2008/01/22/spring-rollback/

130

Common symptoms of problems

Tests that fail intermittently might also point to larger problems in

the code or the infrastructure that it uses, such as unpredictable
background processes or unreliable external resources. In order to
prevent these problems from showing up in production, we may need
to redesign the system to make it more reliable. For example, instead
of directly sending requests to an unreliable external system, we might
batch up requests in a queue and make sure that they arrive at their
destination. As a nice side-effect of this redesign, the system will
become easier to test as well.

Parts of tests or even complete tests used as

set-up for other tests
It is not at all unusual to define an example as an extension or
continuation of a previous example. Sometimes it's a special case,
sometimes it's an alternative workflow. People often create tests for
such examples by copying a part of the previous test, or the whole
test, and using it as a set-up for the new test case. This has to be done
with care because it is the number one cause of bloated tests with lots
of irrelevant information. More often than not, we do not need all
the details of the previous test case in the new one, especially if the
whole test with the trigger and verification is just copied and extended.
The new test will suffer from all the same problems described in the
section about long tests.
A much better solution is to copy only the details that are really
needed, or to extract the preparation phase into a common set-up for
the old and the new test.

Interdependent tests
Tests that depend on the order of execution are just a special case of
the previous problem though this is not as visually obvious as in
the previous case and they suffer from the same problems. If a test
requires some other test to be executed beforehand to prepare the
data or initialise some external dependencies, then a change in the
first test might cause the second to start failing with no obvious reason.

131

Facilitating change

In addition to this, we will never be able to execute just the dependent

test on its own ever, and complex dependencies like this quickly get
out of hand. We need to run individual tests and get quick feedback,
not run the entire test suite every time.

Cleaning up tests
Acceptance tests have to be maintained throughout the software lifecycle. Knowledge gained during development often provides new
insight into the domain and helps to implement, explain or define
things better. This knowledge should be incorporated into existing
tests in the following iterations. Here are a few tips for housekeeping
that you can apply to keep tests flexible and easy to change.

Extract interdependencies into a common set-up

Focus each test on a single business constraint or rule
Strip tests to basics
Convert scripts to specifications
Evolve the language consistently
Organise tests so that they are easy to find.

Extract interdependencies into a common set-up

Tests should be independent. Every test must be able to be run on its
own. This does not mean that we have to introduce a lot of duplication
into the tests. Automation tools often provide a way to specify a
common set-up and tear-down phase for a set of tests and this is how
test dependencies should be resolved. This will allow us to execute
tests independently or together and still change common items at a
single place when we need to.

Focus each test on a single business rule

Make sure that every test checks one and only one business constraint
or business rule. When the functionality changes, we will have to
update only a single test and we will know exactly what we are

132

Cleaning up tests

updating. This practice also helps us pinpoint problems when a test

starts failing.

Strip tests to basics

Clean up tests by removing anything that is not directly related to the
thing described in the test. This will make tests easier to read and
understand. Test stripped to basics are also much more stable, because
they depend only on the minimal set of domain elements. Changes
in other parts of the system will not affect them.
Reducing complexity is also important because we don't want to have
to test our tests. Acceptance tests should be easy to comprehend and
verify. If they get to the point where it is no longer obvious what the
example in the test specifies, then the test loses value as a communication device. I have stressed this several times in the book but I'll do
it again: the real value of acceptance tests is in communication. This
is not to say that more complex test scenarios should not be written
or executed, but they do not belong in the acceptance testing suite.
An example of this would be a script that verifies integration with a
third-party credit card processor by running a set of a dozen cases
with each of the fifty test credit card numbers. Such a test script could
be described by a table with ten columns (input arguments and
expected outcome) and hundreds of rows (different combinations of
arguments). This is a very useful integration or regression test, but
it is not easy to read or communicate, so it is far from ideal as an
integration test. From the perspective of describing what the system
does, the same could be specified with a much smaller and more
focused test script that is easier to understand.

Convert scripts to specifications

If you notice a number of similar tests, this is most likely a sign that
you are missing an important abstraction or trying to describe how
the system does something instead of directly specifying what it does.
A common symptom of this is describing business rules through a
series of workflows. Look for the underlying rules, constraints and
decisions in the group of similar tests and replace all of them by a
133

Facilitating change

single specification focused on the underlying functionality. Start by

identifying parameters and values that change in the tests. From them
identify those that are really important for the goal of the test and
remove the others to make the specification more focused. See the
section Distilling the specifications on page 78 and the section Keep
tests human-readable on page 87 for more details.

Evolve the language consistently

If you rename classes and objects in code, make sure to propagate the
changes to the acceptance tests. Refactoring can make the code
language slowly drift away from the test language, creating ambiguity
and confusion later on. There must be no technical and business
jargon on the project, only the project's ubiquitous language.
Some test automation tools allow you to bind the test descriptions
directly to your domain objects, making it easier to spot language
inconsistencies since tests will start failing after a change on only one
side. In any case, a change in the domain language should be agreed
and communicated to everyone involved developers should not be
changing the language on their own.

Organise tests so that they are easy to find

Because acceptance tests are primarily a communication device, it
must be relatively easy to locate a particular example and similar or
related examples. Almost all acceptance testing tools today allow us
to categorise and organise tests into hierarchies and group related
tests. There is no universal solution or strategy for this and you should
work out what makes most sense for your team and environment,
but I want to stress that tests that are not efficiently organised quickly
get out of hand.
One way to categorise tests is by iteration, user stories in an iteration
and tasks for a particular story. Another is by application module,
business rule and then variants of the rule. The first strategy makes
more sense for tests in development, as it groups together the target
of an iteration while people are working on it. The second strategy

134

Keep tests in a version control system

makes more sense for long-term test archiving, because it makes it

easy to identify and locate tests by business functionality or area of
influence. The best approach is to use the first strategy for tests as
specifications in development and the second strategy for tests as a
record of the system (and regression suite) after development. This
means you need to reorganise tests after the iteration ends.
Some acceptance testing tools allow you to set up links in the hierarchy, so that the same test appears effectively in two places but you
need to update it only once. With a tool like this, you can actually
have both strategies for test organisation in parallel.

Keep tests in a version control system

Acceptance tests define what the code should do and explain what
the current code does. They should live (and die) next to the related
code. This specifically means that acceptance tests should be kept in
the same version control system with the same version organisation
as the domain code. A good feature of most acceptance testing tools
today is that they store tests in a plain text or HTML format, which
works really nicely with source code control systems.
Even if the tool that you choose for acceptance testing allows you to
keep the tests separately from the code, resist this urge. It might not
be very intuitive, but storing acceptance tests separately from the code
will cause huge consistency problems when new versions are introduced or existing code branches are merged. Remember that acceptance tests are a live specification. If you keep the specification and
the implementation separate, they will evolve separately. It will not
be easy to determine which version of the code relates to the
specification, or how to verify a particular version or branch of the
code. When the code and the tests are kept in the same version
structure, under the same version control system, then the connection
between the two is obvious.
Keeping tests in the same version structure as the code also helps if
you need to resurrect parts of the system that were modified or deleted
by mistake. Recovering both the specification and the implementation

135

Facilitating change

at the same time is then easy, as is verifying that they are still
consistent.

Stuff to remember
Acceptance tests are an authoritative description of what
the system does.
We can have the same level of confidence in acceptance
tests as in executable code.
We can use existing acceptance tests to discuss future
changes.
Once the functionality is implemented, previous tests
should always pass.
If an earlier test fails, immediately discuss with customers
whether it specifies obsolete functionality. If yes, take it
out. If not, you found a bug.
Regression tests in doubt should never be disabled.
Automate periodic execution of regression tests using a
continuous integration tool.
Acceptance tests should not be the complete regression
testing suite. They are only a good start.
If you find yourself resisting changing code in order not
to have to fix tests, you need to simplify the tests and
make them easier to maintain.
Efficient organisation of tests is crucial to keeping them
easy to maintain.
Keep tests in the same version control repository as
domain code.
Watch out for these problems:
Long tests, especially those that check several rules
Parameters of calculation tests that always have the
same value
Similar tests with minor differences

136

Keep tests in a version control system

Tests that reflect the way code was written or tests that
mimic code
Tests that fail intermittently even though you haven't
changed the code
Parts of tests or even complete tests used as set-up for
other tests
Interdependent tests

137

138

Part III. Implementing agile

acceptance testing
In this part, we take a look at some more practical aspects of agile
acceptance testing, how it fits into the wider project plan and individual
iterations and how to implement it in an organisation. I also briefly
present the most popular test automation and management tools and
suggest what you can expect from them in the future.

Chapter 8

Starting with agile

acceptance testing
Agile acceptance testing is, primarily, a communication and collaboration technique. It introduces a big change in the way that the
members of a software project talk to each other, transfer knowledge,
specify and document their project. I believe that starting to use agile
acceptance testing is more an organisational change than a technical
change. It affects the traditional roles and responsibilities of
developers, business analysts and testers. The tools and technologies
that they use are of much less importance for the change than the way
teams are organised.

Agile acceptance testing in the

context of the development process
Agile acceptance testing is not a development methodology. It tells
us what we need to deliver, not how to actually develop and deliver
software or how to plan the project. Every now and then someone
tries to sell the idea of acceptance test-driven development methodology on the Internet, but I really do not buy into this. In fact, there
are so many of these X-driven-Y names now that I am really suspicious
of anything new that comes in this form. Agile acceptance testing is
a communication practice and it helps us build a shared understanding
of the domain. The acceptance criteria constititute a target for
development, but they do not prescribe how to actually develop the
system.

141

Starting with agile acceptance testing

Jim Shore gave one of the best summaries of agile acceptance testing
in the wider development context in his article How I use FIT.1 He
called the process describe-demonstrate-develop.
1. The first step in the process, describe, requires us to say what we
are going to develop using a short description, no longer than a
paragraph of text.
2. The second step, demonstrate, asks us to show various differences
in possibilities with examples and captures the actual specification
for the work.
3. The third step, develop, is the implementation of the specification
using regular development practices and methodologies.
4. The fourth step, although it does not appear in the name of the
process, is repeat. Each pass through the describe-demonstratedevelop cycle should elaborate one small business rule. A project
might consist of hundreds or thousands of such cycles.
Agile acceptance testing has much more to say about what happens
before development than during it. John von Neumann, the father of
modern computing, said: There is no sense being exact about
something if you don't even know what you're talking about. The
first two steps make sure that we know what we are talking about
before the real development starts.
From a theoretical perspective, the approach of using realistic
examples for discussion and learning about the domain and then
selecting a set of them as acceptance tests works with any development
system or methodology. In practice, however, the approach works
best in the context of agile development methods (hence the word
agile in the name). I have not tried it in a project that uses the waterfall-style method and honestly I have no intention of trying it in that
environment, but my gut feeling is that it would not work so well
there. Agile methods break up the project into small time-boxed
iterations. Each iteration aims to implement a relatively small, easily
manageable piece of the total project scope. Specification workshops
to discuss the next two weeks or month of development can be held
1

http://jamesshore.com/Blog/How-I-Use-Fit.html

142

Agile acceptance testing in the context of the development process

efficiently, often in one afternoon or a working day. This means that

workshop participants do not spend too much time away from their
normal duties, so it is much easier to get stakeholders to participate.
The small amount of scope means also that the set of tests that come
out of the workshop are relatively easily and quickly reviewed, if signoff is necessary.
With development methodologies that require specifying and
discussing large amounts of scope up front, these workshops would
last for much longer. Holding up so many people for so long at a time
would probably not be feasible. Instead of decision makers and
domain experts we would get junior representatives whose time is
less valuable. The quality of their feedback would be much lower than
the insight of senior domain specialists. With an agile development
method, instead of one huge workshop that lasts a month we can have
thirty one-day workshops during the project. Getting domain experts
away from their daily duties for a day or an afternoon at a time is
much easier than getting them for a month.
Agile development methodologies allow us to adjust and change
requirements and specifications on the fly, which means that if
someone discovers a gap in functionality, we can just easily add
another test without really upsetting anyone (as long as they are told
that the specification has changed). Having this flexibility is very
important, especially when the team is learning how to use agile
acceptance testing, since the cost of mistakes in the first two stages
of the process is very low and we can allow ourselves to experiment.
The incremental and iterative way of developing also allows us to
incorporate the knowledge learned during the project into our
examples, specifications and acceptance tests. In the first few workshops the developers may not know anything about the domain, but
by the fifth or sixth workshop they will know quite a lot and be able
to contribute a lot more. If we only had one big workshop before the
development started, then the feedback that developers bring to the
table will not be anywhere near as valuable.

143

Starting with agile acceptance testing

Agile methodologies also do not require a lot of paperwork, so you

can build up the specification on the fly using examples from agile
acceptance testing. Acceptance tests are the specification of the system.
As we add acceptance tests and evolve existing ones, we build and
refine the specification. This saves a lot of time and effort because
you do not have to produce and maintain large requirements and
specifications documents. In fact, having both these documents and
acceptance tests might introduce confusion if they are not perfectly
in sync. This benefit is obviously lost with methodologies that require
sign-off on formal plans and documents (although, as suggested in
Chapter 5, you can get sign-off against sets of acceptance tests).
This does not mean that you cannot use agile acceptance testing if
you are not agile by the book. People are often scared when too many
buzzwords arrive together, so for those of you that are thinking about
migrating to agile, I'd like to point out that you do not have to eat the
whole elephant at once (or ever finish that meal). Doing unit testing
or pair programming or any other popular agile practice is not a
prerequisite for agile acceptance testing, and you definitely do not
have to learn all these ideas and apply them at once. Agile programming techniques certainly help, not because agile acceptance testing
somehow specially depends on them, but because they are generally
useful practices. An iterative development model is, however, a
prerequisite for effective agile acceptance testing.

Fitting into iterations

The idea of specification workshops that support acceptance testing
seems to cause a lot of confusion and even misunderstanding. In many
public talks I give about agile acceptance testing, I get questions from
people who somehow contrast it to iterative development and mistake
the workshop for a big design session upfront, expecting that it will
lengthen the feedback loop. Agile acceptance testing does not say
anything about how to develop software or organise iterations. There
is nothing stopping us from fitting it into an iteration. In fact, it works
best with short iterations. Short iterations force us to keep the stories

144

Fitting into iterations

discussed in the workshop short, enabling us to keep the workshop

short and focused.
Here is an example of how the specification workshops (and acceptance testing) fit into an agile process to shorten the feedback loop
and improve iterative development. The diagram in Figure 8.1 is a
version of the iteration workflow that Andy Hatoum and I came up
with while I was writing this book. Take this as just a guideline, not
as something that you must follow exactly. Hopefully it will make it
easier to understand how to apply these ideas in different environments. The contents of the brackets indicate the iteration that a step
relates to: N is the current iteration, N-1 is the previous one and N+1
is the next.

Figure 8.1. A suggestion for iteration workflow

Release (N-1): If we wish to, we can release the result of a previous

iteration on Monday. Not every iteration has to be shipped to
customers, but we like to have the code as a potentially shippable
software package at any point. We don't release on Fridays to reduce
the risk of causing production problems over weekends while people
are not readily available to fix issues.

145

Starting with agile acceptance testing

Implementation (N): On Monday the implementation of the current

iteration starts. This includes all normal agile development practices
such as unit testing and continuous integration. An important change
for teams already practising agile development is that the development
is focused on implementing acceptance tests specified for the iteration.
Implementation includes running and cleaning up acceptance tests,
identifying and discussing missing cases, writing more tests for these
cases or modifying existing tests. It also includes releasing to the QA
environment often, so that on-site customers and testers can use the
software and provide quick feedback. Implementation continues until
Wednesday of the second week.
Pre-planning (N+1):The clients, the business analyst, the project
manager and the development team leader work out roughly what
will go into the next iteration (not the current one, but the one after).
This is intentionally a rough plan, not carved in stone, but good
enough to give the business analyst and the customers a starting point
for discussion and allow them to keep one step ahead. This happens
on Monday as well.
Acceptance test clean-up and review (N): Not all acceptance tests that
we have prepared for this iteration will be perfect straight away, so a
business analyst chases open questions and gets remote domain
specialists to review the tests, while the project manager can get any
sort of sign-off required for tests, and testers, developers or business
analysts simplify, clean up and organise acceptance tests. We expect
this activity to span the first few days of an iteration. Changes might
be introduced into the acceptance tests later, but we expect the bulk
of the work to be done in the first few days and for the tests to stabilise
after this. The diagram shows that this should end by Thursday, but
this is not a fixed deadline, it is an estimate. Clean-up ends when it
ends, sometimes it may not be needed at all, sometimes it will end
sooner, sometimes it will spill over to the next week.
Preparing examples (N+1): Once the bulk of the acceptance test cleanup for the current iteration is done, the business analyst can start
working with the clients and the testers on the examples for the next
iteration, preparing a good starting point for the workshop. The

146

Fitting into iterations

deadline for this is again an estimate. In general, after the work on

the current iteration is done, the business analyst should focus on the
next one and stay on it until the polishing on release N starts. The
important thing is that the examples don't need to be absolutely
complete at this point. The goal is to get started and prepare for the
workshop, not to iron out all the acceptance tests. This is explained
in more detail in the section Keeping one step ahead on page 61.
Integration testing, Exploratory testing, Polishing and Demo (N):
The development work should be done by Tuesday in the second week
of the iteration. At this point, we start polishing the result. The
developers will already have been releasing builds to the internal QA
environment during implementation, as soon as there was something
for others to see. This allows the on-site client and testers to check
what developers are doing during the development and provide
feedback, but these intermediate releases do not necessarily need to
be production quality. However, at this point in the cycle we start to
really focus on producing a clean, polished package. This involves
cleaning up database scripts, packages, configuration, upgrades from
the previous version and testing the installation. Everyone switches
from adding new features into exploratory testing and cleaning up
the results. Issues, UI bugs and ideas for small improvements get
resolved during Wednesday and Thursday. By Thursday afternoon,
we should have a polished version that we can demonstrate to off-site
customers remotely. Polishing may spill over on to Friday, but this
should be an exception rather than a rule.
Retrospective (N): This is a normal retrospective, normally looking
at the last two weeks of development but occasionally a longer period.
Everyone is involved.
Specification workshop (N+1): We select stories for the next iteration
and then nail down the specifications during the workshop by
discussing relevant examples. The business analyst introduces the
examples that she has prepared with the customers, then we all go
through them. The other workshop participants ask questions and
suggest edge cases and new examples for discussion. Developers need
to think about how they will implement the stories and identify

147

Starting with agile acceptance testing

functional gaps and inconsistencies. Testers need to think about

breaking the system and suggest important testing cases for discussion.
We can organise feedback exercises during the workshop to check
whether we all understand the same thing. The aim of the workshop
is to build a shared understanding among developers, business people
and testers about the aims for the next two weeks of work. Because
realistic examples are discussed and written down, inconsistencies
and gaps should be easy to identify and we should get a solid foundation for development. The workshop ends when everyone involved
agrees that there are enough examples and that everything is clear
enough to start the work. Some questions may remain open, because
they require the approval of a more senior stakeholder or the opinion
of a particular domain expert, but in general the workshop should
define the precise scope of the next two weeks' work and ensure that
we all agree on this. A representative set of examples is chosen to be
an acceptance criterion, and formalised into acceptance tests.
Planning (N+1): Once everyone knows what the potential scope is,
we organise estimations, play planning poker and confirm the final
scope for the next iteration. This becomes the official plan. Everyone
is involved.
After this, we start over with the next iteration.

Igniting the spark

Introducing agile acceptance testing in an organisation by decree and
without real buy-in from the people affected will only meet with
resistance and cause the attempt to fail. I think that a much better
way to introduce agile acceptance testing in an organisation is to start
out small and then let this spark ignite a larger change.
The best thing, in my opinion, is to try out agile acceptance testing
with a team of people enthusiastic about the experiment on a midsize project or a module of a larger project. Small projects do not have
the communication challenges that would demonstrate the advantages
of agile acceptance testing and they are too short for best practices to

148

Igniting the spark

evolve. A mid-size project can be completed with a relatively small

team so you should have no problem finding enough enthusiastic
people to try out agile acceptance testing, yet the team will be large
enough to demonstrate how agile acceptance testing as a practice
helps to improve communication and collaboration. In an organisation with traditional roles, you probably need one business analyst,
a few testers and a few developers on the team. It is important for this
initial team to be composed of enthusiastic people because they will
look for ways to overcome difficulties and adjust general ideas to their
environment, rather than just abandoning the new practice and falling
back to the usual way of working at the first sign of trouble. Have no
doubt, this kind of organisational change will face obstacles on the
way. Although most people get into agile acceptance testing within a
few weeks, it might take a few months for the ideas to fully settle down
and the full benefits of the practice to show.
Once this project starts delivering, the ideas will simply spread virally
through the organisation. Boosted by fresh ideas on how to improve
communication and collaboration, business analysts will start talking
to other business analysts, testers with other testers and developers
with other developers in your organisation, spreading the word.
Management will notice that there is something different about this
particular project because suddenly there are fewer problems in
development. People from other teams will phone you to ask how
this testing thing works.
As with any agile software practice, nothing should be taken as carved
in stone and the initial team members should work out the best ways
of applying the ideas of agile acceptance testing in their own environment. These can then serve as much better guidelines for rolling out
the practice on a larger scale than the general ideas. You can then
swap the people on the original team with some people on other teams
and use them as trainers for a larger roll-out.

149

Starting with agile acceptance testing

Don't mention testing

Developers are often enthusiastic about agile acceptance testing.
Testers typically sit on the fence, with testing managers sometimes
openly fighting against it because they fear that someone is trespassing
on their territory. But it's business analysts in a corporate environment
who often put up the greatest resistance to agile acceptance testing.
If you want to introduce the practice in a corporate environment with
strictly defined roles, avoid using the word testing at all. Testing in
the name of the practice often causes a lot of confusion and resistance.
Talk about executable specifications or communicating with examples
instead of acceptance tests. (See the section Better names on page 38
for more ideas.) I've used this approach to deal with business analysts
who object to the whole practice because they don't do tests and with
testing managers that object to someone else testing the software.
You can organise a specification workshop and formalise the examples
that will be used as a specification without ever mentioning that they
will be used for testing. Ask business analysts to be there to help
programmers work with a new experimental way of building the
development specification. Ask testers to be there to provide additional feedback and to learn about the domain. Ask developers to
participate to flush out inconsistencies from requirements before the
development starts. Then just convert the specifications to tests
without involving anyone who might object to the whole concept of
automating acceptance tests for development. You can include people
who genuinely want to help, be they testers, developers or business
analysts, but the rest of the world does not necessarily need to know
that you are doing anything related to testing. This is not an ideal
situation, but it will be good enough to start with. Make sure to verify
the tests with business people once they are written down, again
avoiding mentioning the word test. Just ask whether you have written
down the meeting minutes correctly. When the time comes to discuss
the ideas again, instead of the original Word document, show people
FitNesse or whatever tool you are using and just tell them that this is
how you decided to store the specifications. When the time comes to

150

Adopting in phases

test, offer testers a set of already automated verifications for some of

the test cases and work with them to define the rest.
After a while, the benefits of this practice should become obvious and
then you should be able to tell everyone the full story and start using
the name acceptance tests.

Adopting in phases
Brian Marick wrote an article called An alternative to business-facing
TDD2 in March 2008, challenging the effectiveness of automation for
acceptance tests. Arguing that test automation by itself does not help
clarify the design, that live demonstrations can show continuous
progress and that automated acceptance tests cannot detect user
interface bugs, he concluded that test automation for business facing
(acceptance) tests does not pay off as much for code-oriented (unit)
tests. He suggested that it might be more effective not to convert the
examples into automated tests and perform exploratory testing for
acceptance:
An application built with programmer TDD, whiteboard-style
and example-heavy business-facing design, exploratory testing
of its visible workings, and some small set of automated
whole-system sanity tests will be cheaper to develop and no
worse in quality than one that differs in having minimal
exploratory testing, done through the GUI, plus a full set of
business-facing TDD tests derived from the example-heavy
design.
Although I do not completely agree with this hypothesis, I think that
it is important to include it because Brian Marick is one of the top
authorities on testing in agile projects and definitely has much more
experience with acceptance testing than I do. His article is relatively
short and is a very good read, so if you are sitting near a computer,
put this book down and take five minutes to go through it now. If

http://www.exampler.com/blog/2008/03/23/an-alternative-to-business-facing-tdd/

151

Starting with agile acceptance testing

you are not close to a computer, make sure you remember to read it
later.
I agree that automated acceptance testing can never replace exploratory testing for finding UI bugs and identifying unforeseen problems,
but I do not think that these two practices should be competing against
each other. With today's automation tools, user interface testing is
still a big pain and I consider that exploratory testing is much more
effective for user interfaces than automated testing. On the other
hand, I see a lot of value in incrementally building up an automated
regression test suite for business rules. Automated acceptance tests
not only provide assurance that the software meets the goals, they
also serve as a very good safety net during design improvements and
changes (see the section Evolving the design on page 109). They also
leave more time for exploratory testing.
The idea of stopping after the specification workshop is interesting
from another aspect. The workshop helps to build a shared understanding of the domain and working with realistic examples helps us
get more complete and precise specifications. This is, for me, the
biggest benefit of agile acceptance testing at the moment. If you think
that introducing the whole practice at once might be too much for
your team, start by only doing the workshop and then slowly move
into test automation.

Choose a facilitator for the

workshops
A very good way of making specifications workshops more effective
is to assign a facilitator. The role of the facilitator is to ensure that
the workshop is on track and that it does not turn into yet another
boring meeting. The facilitator should make sure that everyone is
participating, both by stopping people from taking over the workshop
and by encouraging (even directly requesting) feedback from silent
participants. He or she should set the topic for discussion and have
the right to change it when the discussion gets sidetracked, especially

152

Choose a facilitator for the workshops

watching out for personal bickering and parallel discussions. If a

developer or tester is trying to explain why an edge case is important
and the other participants are dismissive, the facilitator should help
the person formulate the example better. If, however, business people
acknowledge and understand the edge case and they still dismiss it,
the facilitator should ensure that the discussion returns to the cases
that are relevant for this particular phase of the project. The facilitator
is also responsible for making sure that people discuss realistic
examples at an appropriate level of abstraction and in enough detail,
preventing the discussion from digressing into unrealistic contrived
examples or implementation specifics, but ensuring that it still
provides enough context to build a deep shared understanding.
The best person to be a facilitator is someone who has the right people
skills and an overall understanding of the domain, possibly not in
full detail. This person should be able to speak both to technical and
to business people, and help them find a common language. If the
team includes a full-time business analyst, she is probably the best
person to facilitate the workshop. The facilitator does not have to be
a domain specialist. A ScrumMaster or a team leader might also be a
good choice. By team leader, I mean someone with personal leadership
skills who spends a lot of time communicating with people, not
someone who has the role because of great technical knowledge. A
key thing to remember is that instead of translating between business
and technology the facilitator must help other workshop participants
to find a common language.
Writing about collaborative requirements workshops, Ellen Gottesdiener insists that the workshop needs to be a facilitated meeting,
going as far as advising against using a workshop for requirements
unless you have a facilitator who is neutral to the outcome ([19] Ch.1).
Although the specification workshop can be considered as a kind of
requirements workshop, Im not so sure about this prerequisite of a
neutral facilitator. I can see the value of a facilitator when a new team
is learning how to apply agile acceptance testing, but I believe that
smart people can be self-organising once they know what they should
be doing. I have often played the role of a facilitator, but I was very
much involved in development and never neutral to the outcome.

153

Starting with agile acceptance testing

Even so these workshops produced good results. A good facilitator

does not necessarily need to be neutral to the outcome, but the person
should be able to take a neutral point of view or even play the devil's
advocate to evaluate and stimulate the discussion.
An alternative to relying on facilitators long-term is to promote the
culture of a good fight, where open conflict of opinions is encouraged
and respected rather than avoided. Naturally, this good fight concept
should be limited to the topic at hand and not diverge into personal
insults. My opinion is that a facilitator, although certainly helpful, is
not essential if the team knows and honours the rules of the game
and the goals of the workshop. While the team is still learning what
the workshop should be about and defining the rules, a facilitator can
help significantly.

Hire a mentor
It always helps to have someone who has gone through the transition
to lead the way, answer questions and train project team members.
If you can hire someone like this and get him on the initial team, this
will help to establish best practices for acceptance testing in your
organisation much more quickly. The mentor can act as a facilitator
during workshops, help to choose tools, train people to use them and
help automate tests better.
Having someone readily available to answer questions and help with
problems will speed up progress, but it might also make the team too
reliant on the mentor. The ultimate goal is to empower your initial
team to work without external help and then even train other teams
in your organisation. I think that an especially effective technique is
to have someone to help out for a short period of time, such as a few
weeks or a month, and then have them go away and let the team
members work on their own for a few weeks. Once the mentor goes
away, the team members will have to try harder to solve issues and
problems on their own, so they will learn more from these challenges.
If you work on shorter iterations, then the mentor should be with the
team during the initial few iterations, then leave the team to work

154

Where to look for a mentor

out examples and produce tests for one iteration themselves, joining
towards the end of it to review the outcome and help them solve any
open issues before the specification workshop for the next iteration.
As the project progresses, the mentor should join for shorter periods
of time and leave for longer, making the team more and more selfreliant. After a few such visits, the mentor can just occasionally visit
the team to discuss progress. He can be available for phone or e-mail
consultations throughout. A good mentor will pass the knowledge
on to the initial team in a few visits, so the entire experiment should
not take more than two or three months.

Where to look for a mentor

Agile acceptance testing is an emerging practice and there are no
official organisations or professional societies that you can turn to
and ask for a mentor. I would be wary of any consultants who try to
sell themselves as certified trainers for this practice. This is not to say
that someone with a certificate is necessarily a fraud, but look for
experience rather than paper qualifications. Your best bet for the role
of mentor or trainer would be someone who has already gone through
this transition in another company similar to yours, so you could try
to snatch someone away from the competition.
Don't look for a mentor in an official organisation look for a
consultant on mailing lists such as Agile Testing (agile-testing), Agile
Aliance Functional Testing Tools (aa-ftt) and FitNesse Yahoo groups
and the Agile Acceptance Testing group on Google groups (links are
in the Resources appendix). My company provides in-house and
public training courses for agile acceptance testing. See
http://neuri.co.uk for more information on this.

Avoid the cargo cult mentality

In some places, developers try to get acceptance testing introduced
from the bottom up by writing acceptance tests themselves, using the
requirements documents as a guide, without discussing the tests with

155

Starting with agile acceptance testing

business people. This might seem like a good start, but it is actually
as wrong as it can possibly be.
Cargo cults
During World War II, Allied forces and Japan invaded isolated
islands in the Pacific where the natives had had no contact with
the rest of the world before. They arrived with modern equipment, built airfields and delivered supplies by cargo drops. Some
of the supplies were shared with natives, drastically changing
their way of life. When the war ended, the soldiers and their
supplies were gone. Religious cults developed on the islands
calling for the cargo to start falling from the sky again. Islanders
built elaborate models of airplanes and air control towers out
of straw and wood and imitated soldiers, hoping to summon
the presents from the gods. Needless to say the presents never
arrived.
The term cargo cult was introduced to programming by Steve
McConnell, who characterised the ritual following of practices
that serve no real purpose as cargo-cult software engineering
in 2000.3 Since then, this name has became synonymous with
any blind use of programming practices without really understanding the underlying principles, which often does not bring
any benefits.
If developers write acceptance tests based on their understanding of
the system, they are not actually checking whether this is the same as
the understanding of business people. The result is that people are
disappointed because acceptance testing seems not to provide any
benefits apart from regression test coverage. If you think about it,
since there was no effort to build a shared understanding, there can
be no benefits of improved communication. Acceptance tests written
in this way are no better than the straw air-control towers and wooden
planes that the Micronesian natives built. They follow the outer
3

[20] and http://stevemcconnell.com/ieeesoftware/eic10.htm

156

Avoid the cargo cult mentality

structure and provide an image of something useful, but the core is

missing and the benefits are simply not there.
If developers decide that they want to implement acceptance testing
and there is no buy-in from the business experts to organise
specification workshops, there can still be some benefits but they will
not be as effective as with full cooperation from the business experts.
A model that may work is for developers to write tests after the
requirements are handed over but before development starts. They
can then ask the business people to verify the tests, trying to initiate
a discussion during the verification. This approach can be implemented on a small scale, with one or two developers discussing tests with
a business analyst or a customer. It will at least allow these people to
synchronise their understanding and might be the first step toward
introducing the full workshop practice later in the project.
In the introduction, I said that you should use the process described
in this book as a guide rather than something carved in stone. You
will probably want to modify some parts to make it fit your organisation better, but make sure that you follow the core ideas and underlying principles. Otherwise you will end up following a cargo cult.

Stuff to remember
Agile acceptance testing is not a development methodology.
Iterative development is a prerequisite for effective agile
acceptance testing.
Start out small, with a team of enthusiastic people.
Think about assigning a facilitator for the initial workshops.
Don't take practices as carved in stone, adjust them to
your needs but keep to the basic principles.
Avoid mentioning the word test to get the buy-in of
people who think that testing is beneath them.
Developers should not write acceptance tests themselves.

157

Starting with agile acceptance testing

If you cannot get business buy-in at least to review the

tests, don't do acceptance testing you will just waste
time.

158

Chapter 9

Planning with user

stories
In this chapter, we take a look at how agile acceptance testing fits into
into a wider project perspective, specifically in the context of user
stories. User stories are today the most popular technique for planning
agile projects. Although there is nothing peculiar to agile acceptance
testing that requires you to plan with user stories, these two techniques
complement each other incredibly well and produce the best results
when they are used together.
I have intentionally placed this chapter towards the end of the book,
because I wanted to emphasise the communication benefits of
acceptance testing and avoid focusing too much on general software
development practices and techniques. How you plan your project
depends very much on the type and size of the project and the period
available to you to develop it, so please don't interpret this chapter
as something that is carved in stone or prescribed as the only way to
do things. You do not have to base your plans on user stories, although
there are significant advantages to using them if you are introducing
acceptance testing. If you understand the underlying principles and
ideas of user stories you will be able to see more easily how to fit agile
acceptance testing into your planning process, and you might even
find that you can improve it.
This chapter is intentionally short. It will probably leave you with
many questions about the practical application of user stories. I want
to encourage you to investigate this topic further. The bible for user
stories is Mike Cohn's book User Stories Applied for Software Development[11]. I strongly recommend reading this book if you have not
already done so. For ideas on how to work with user stories more

159

Planning with user stories

efficiently, also see Agile Estimating and Planning[12] by the same

author.

User stories in a nutshell

A user story is a very brief description of what the system does for its
users to help them perform their work more easily. It describes the
system from the viewpoint of the user, not from a technical
perspective. There are no details of how the system implements the
required functionality in a user story, or even what the detailed
requirements are. All specifics are left for later discussion. Stories are
used to plan out the project broadly, by focusing on customer benefits.
Typical stories are fairly high-level, allowing us to discuss plans and
prioritise deliverables without getting tied up in implementation
details or flushing out all the requirements.
From a more traditional point of view, user stories describe the scope
of a project, but not the specification. At a very high level, they answer
what and why. There are no details because we want to avoid paralysis
by analysis, getting lost in the details and not seeing the wood for the
trees. Leaving the details for later also allows us to keep the initial
scoping discussion short and efficiently use the time of project sponsors and key stake-holders who need to be involved in building a highlevel delivery plan and scope.
I like to use one-liners as user stories; any more is probably too much
detail. The classic template for user stories is:
1. As a role,
2. I want functionality,
3. So that benefit.
The behaviour-driven development community promotes a different
story format, moving the benefit to the first position to emphasise it.
The BDD story template is:
1. In order to benefit

160

How are stories different from more traditional methods?

2. As a role
3. I want functionality
For me, there is no significant difference between these two formats,
but some people religiously insist that one is better than the other, as
if a change in the order could really make an important difference
for the project. This reminds me of the famous curly brace discussions
and similar coding format issues, which were always so pointless but
wasted so much time. Ken Arnold's legendary article1 asking for style
wars to stop always comes to my mind in these situations.

How are stories different from more

traditional methods?
The most important difference between user stories and more traditional techniques, such as use cases, is that user stories are there to
facilitate the discussion and support planning rather than to be an
authoritative description of system requirements. Use cases often live
on throughout the project as a load of printed material with formal
sign-offs, but user stories are just placeholders for discussion and
information which will come later. For much more information on
the differences between user stories and use cases, see [11] Ch. 12.
The worst case of people not understanding the point of user stories
that I have seen was a project plan that included product CRUD
(Create, Read, Update, Delete, typical data entry screen capabilities)
as a user story. A user story might include product maintenance, but
it should also specify for what purpose and by whom the maintenance
is performed and it should describe something that will deliver some
real business value to the customers. When you understand who needs
to create, update and delete products and why, the scope of the story
is often a lot less than in a generic case. A generic product CRUD case
would typically include everything required for product maintenance
in the whole system. A true user story includes only what is needed
1

http://www.artima.com/weblogs/viewpost.jsp?thread=74230

161

Planning with user stories

to provide a particular business benefit to a particular user. Some of

the fields, relationships and validations do not need to be implemented
to fulfil the job of a particular user for a particular story. Even more
importantly, some other functionality might be required to complete
the story and make it really usable in production. For example, if the
products were also presented to customers on a public web site, even
if they were not sold in the first iteration, this might bring some
business value. If creating and updating products is not enough to
bring real business value, then the implementation of a product CRUD
story will just sit in the testing system without actually being used.

Benefits of user stories

User stories empower development teams and customers to prioritise
and scope projects so that customers get real value for money much
faster and the time and effort of everyone involved in project planning
is used efficiently.
I do not think that there is any specific benefit to user stories over
more traditional planning methods if you want to deliver everything
in one big chunk, you are sure that the requirements and the
specifications will not change during the project and your developers
are domain experts. But most of my projects are not like that. Plans
change, requirements are thrown in or out and clients want staged
deliveries. Developers often make wrong assumptions about the
problem domain and misunderstand requirements and they learn
about the domain while developing the software. If your projects are
like this, you will find that user stories use your time efficiently, reduce
waste and provide a very good framework for discussing, planning
and prioritising with key stakeholders. They also help to flush out
wrong assumptions and help developers make better decisions.
User stories are lightweight and, as the name suggests, completely
focused on the users' point of view. Because of this, customers can
easily relate to the stories and have a much better grasp of what the
functionality described by a story actually brings. This makes
comparing, prioritising and scheduling much easier. Instead of asking

162

Benefits of user stories

the client to decide between implementing an e-mail service or

building a graph-optimisation subsystem, we can discuss priorities
in terms of something that they can relate to, such as finding the best
route for travelling or sending confirmation letters in the customer's
chosen language. This is a subtle distinction that can take a while to
settle in the mind when you first start to use user stories. The main
idea is to plan in terms of business benefits and customer experience,
not technical tasks or infrastructure requirements.
Of course, technical and infrastructure constraints may influence
planning and they should do so. For example, it might make a lot of
sense to implement an accounts system before developing an interface
to payment gateways, or to implement support for generic e-mail
notifications before providing language-specific e-mail features. Such
considerations need to be taken into account, but talking in terms of
something that the customers can relate to allows us to plan more
efficiently and to get better feedback from them.
With planning methods that focus on technical tasks, functionality
gets translated into work items and the plan is built based on these
tasks. When the plan changes later or when priorities shift, it is not
always easy to see which tasks are required to completely implement
some business functionality so that it can be used in production. User
stories avoid this problem by skipping the step in the middle.
Customer expectations are not translated to technical tasks. Instead,
they are used directly in planning, so that things don't get lost in
translation. We plan using items that clients can directly relate to,
with a clear understanding of how parts of the plan bring business
benefits. When business priorities change, user stories provide a really
useful framework for discussion and prioritisation. Clients should
have no problems deciding which user stories are no longer required,
or which ones should be rushed through to implement a key
requirement.
Another great feature of user stories is that each story brings a clear
immediate customer benefit. If you build your delivery plan with user
stories, each deliverable brings direct value to the customer. If the
plan is built on a purely technical basis, infrastructure components

163

Planning with user stories

often take precedence in development, so the customers cannot really

start using the software until much later in the project. With a plan
based around user stories, it is not uncommon for even the first
deliverable to go into production and be used to solve real problems.
This builds confidence in the product early, avoids big-bang integrations and enables customers to give us much better feedback. In fact,
that's the way I prefer to deliver my projects I simply don't trust
software that is not live.

Three Cs of user stories

Ron Jeffries wrote a famous article on three essential aspects of user
stories called card, conversation, and confirmation.2 These three
aspects later became known as the three Cs of user stories.
Card: User stories should be written on small index cards. The card
allows you to write down only so much information, so nobody
can mistake it for a complete requirement. The card is just a
placeholder for discussion and a tool for planning.
Conversation: the card is used just as a token for discussion,
reminding people what to talk about. The conversation is where
we flush out all the details and get to real requirements.
Confirmation: each user story should be verifiable, and we should
have a clean way of deciding whether it was implemented correctly
and completely.

User story cards

Cards can be shuffled around, grouped with other cards and put on
the wall, helping us to see the big picture and plan efficiently. Because
of this, it makes a lot of sense to put information that helps with
planning, such as priorities and estimates, also on the card. A very
good idea is to choose a concise story title and put this on the card as
well, so that people can easily refer to a particular user story.
2

http://www.xprogramming.com/xpmag/expCardConversationConfirmation.htm

164

Project planning

Connextra pioneered the index story card format that is now more
or less considered standard. The Connextra story card has the story
title and priority at the top, followed by the story description, and
then the story author, submission date and implementation estimate
at the bottom. An example is shown in Figure 9.1.

Figure 9.1. The Connextra card template

Project planning
I like to use a brainstorming session with key stakeholders at the
beginning of a project to create the first cut of user stories, writing
them down on index cards or sheets of paper that are then put up on
a wall. We first discuss very high-level stories (sometimes called epics)
and then break them up into smaller stories later. The way that the
stories are gathered is not really important for the topic of this book
see the recommended books on user stories for some good techniques for that. What is important is that each story should have
something to say about how the system will help users do their job
when it is complete, and why this particular functionality is important.

165

Planning with user stories

During the brainstorming session, it is not uncommon for people to

add some stories that seem to be related to the project, or describe
what they would like to have in the long term, but which turn out to
be less important as the others. Some stories might even conflict with
other stories. Once the first cut of the user stories is captured, key
stakeholders need to decide which ones are actually going to make it
into the project. Having stories displayed on the wall helps everyone
to see how big a particular project is going to be, and the overall
picture helps to persuade people to throw out irrelevant stories and
focus the development. We can use the vertical grouping of the sheets
to signal the benefits or goals of a group of related stories (for example,
group all stories that automate transaction processing together), and
use an abbreviation or some other indication to specify the user role.
So instead of the full description As a security investigator, I want
the system to automatically identify high-risk transactions so that I
do not have to manually analyse them we just write identify highrisk transactions on the sheet of paper and put it on the wall. This
short description later becomes the title for the story. In any case, we
should be able to link a sheet of paper on the wall to the user role and
know why the story is on the wall in the first place.

Grouping into deliverables

Collected together, all selected user stories make up the scope of the
project. Armed with this collection of stories, we can split the project
into deliverables that are complete in a sense that they can be used in
production. This forms the basis of our release plan, which we can
then split into iterations ordered according to priorities. As a rule of
thumb, I suggest grouping user stories into deliverables so that each
deliverable has a clearly defined business goal target. This allows us
to prioritise by focusing on business goals and benefits rather than
on technical dependencies. In a recent project for an e-commerce
system, we split the fraud detection system into three deliverables:
1. The first deliverable automatically retrieved and consolidated fraud
alerts from various sources and opened investigation tickets, saving
the investigators a lot of manual work.

166

Grouping into deliverables

2. The second deliverable automated part of the investigation process

by automatically blocking the disputed account and notifying
external partners that all pending financial operations were to be
frozen. This deliverable automated a time-critical part of the
investigation, so that potential fraudsters could not do any more
damage. The investigators then had more time to look into the
actions that had taken place on a disputed account and decide
whether it was really a fraud or just a false alert.
3. The third deliverable automated the decision process for obviously
fraudulent cases, again saving a lot of time for the investigators
and allowing them to focus more on the cases where a human eye
was needed.
Focusing each deliverable on a clearly defined business goal makes it
is easy to decide whether a story belongs there or not. Each story
clearly states why it is required (... so that), so it is obvious which
business benefits are fulfilled by a story. This is one more benefit of
planning with user stories. Focusing on goals means we can identify
stories that have no related goals and leave them out, making the
project much more coherent and providing grounds for discussions
about why something should not be in scope (or at least not yet).
There is more than one way to skin a cat, so you should try reshuffling
stories and deliverables a few times looking for the plan that brings
customers the most important benefits the fastest and makes sense
from a technical perspective. Once everyone is happy with the plan
for deliverables, a high-level project plan is more or less already in
place it's on the wall! Although there are no details in the stories,
technical experts should be able to give rough estimates for how long
the development will take and how many people they need for the
project based on the list of stories.
A crucial idea, from a planning perspective, is to build blocks of
software that can actually go to production. There are so many reasons
for this that I do not know where to start. Project sponsors will be
really happy when they actually have something to show to justify the
costs of development. Software that saves money or brings in new
business has a much better effect than something you can just

167

Planning with user stories

demonstrate on a testing system. The customers will start getting

business benefits early. End-users will start flushing out small problems and identifying unforeseen functional gaps earlier. Programmers
will receive much better and much earlier feedback on their work,
allowing them to build this knowledge into the later phases of the
project. There are many other reasons why iterative development is
now preferred to the classic waterfall model, and they all apply here.
In essence, we want to go live with software as soon as we can, and
then keep delivering chunks that also go live at regular intervals.
A good delivery plan has these features:
1. Each individual deliverable is relatively small so that it can be
developed and delivered in at most a few months.
2. Each deliverable can actually go live and it will be useful to users.
3. Deliverables will take about the same time to develop, so that we
can release deliverables at similar intervals.
4. The order of development technically makes sense.
I use a simple question as a rule of thumb to decide whether a deliverable is functionally complete: Pretend it's magic and we already
delivered this piece of software. Would it actually be useful and usable
in production? If the answer is No, then we discuss which stories
are missing to enable this.
The user stories method does not give you detailed estimates of how
much something will cost or how many months it will take to develop.
On the other hand, I would argue that even a seemingly precise
traditional method based on detailed requirements would just produce
an educated guess that would not necessarily be closer to the actual
outcome than the suggestion of an experienced team leader. And such
detailed methods take a lot more time and effort from key stakeholders, whereas planning with user stories makes sure that their time is
used efficiently. Senior stakeholders are often quite busy and working
on a number of other projects, so it is really important that they focus
on the plan and give real feedback. In my experience, it is much easier
to get people's full attention for one afternoon than to chase them
around for endless meetings until all the details are flushed out.
168

Changing the plan during a project

Changing the plan during a project

The initial plan is used to scope out upcoming iterations. A definitive
set of stories is selected for an iteration just before it starts, and they
are discussed in detail and precisely scoped out. We leave out features
that will be developed and delivered later. When the current set have
been implemented, the customers or product owners have a chance
to revisit the plan, potentially modifying the priorities and shuffling
things around for the next iteration.
On longer projects, the customer's business requirements or market
circumstances often change during the implementation, so parts of
the original specification become obsolete halfway through the project.
Working out a detailed specification for a twelve-month project before
development is a sure way to waste a lot of time discussing functionality that will become obsolete. Because there are no details in user
stories, we can easily change the plan later without wasting too much
time and effort. We won't have spent weeks working on the details of
features that might become irrelevant in the future. A plan based on
user stories facilitates change rather then impedes it.
In a sense, stories are just reminders for discussion and placeholders
in the project plan. Before the development actually starts, we need
to have a proper discussion about what a story entails and flesh out
the details for development.

User stories and acceptance tests

User stories provide a high level plan and facilitate discussion, but
they do not specify any details. Agile acceptance tests, on the other
hand, provide specification details but do not even try to put things
into a wider perspective. Used together, these two techniques
complement each other.
A key idea of user stories is that we postpone a detailed discussion
until we really need to start working on a particular story, to reduce
the overhead of change. Postponing the discussion about details also

169

Planning with user stories

allows us to utilise the knowledge we gained developing previous

stories. This is a major distinction between user stories and more
traditional planning methods. If the entire scope and requirements
are captured only at the beginning of the project, you cannot use any
of the knowledge or insights that you acquire only during development.
When the time comes to have the detailed discussion, we have a
specification workshop (described in detail in Chapter 4). A workshop
takes place at the start of every iteration, once the stakeholders have
chosen a set of stories that they would like it to cover. Before the
workshop, a story is just a placeholder for discussion. During the
workshop, we take the story and discuss it at length. After the workshop, we have detailed requirements and specifications in the form
of examples, built just in time for the development to start.
The specification workshop, in a sense, continues where the user
stories leave off. User stories summarise ideas about the project and
schedule them. The specification workshop is where we elaborate
these ideas and specify exactly what they entail. User stories drive
each workshop and examples defined during it become the acceptance
criteria for an implementation of a story. The discussion that takes
place during the workshop ensures that customers, domain experts
and the implementation team share an understanding of what the
story requires. The acceptance tests produced based on the examples
coming out of the workshop allow developers to focus their work and
testers to verify the story implementation.
Mary Poppendieck says that the worst waste of software development
is adding features that are not needed to get the customers' job
done[21]. User stories, with their focus on customer benefits and user
roles that actually benefit from them, give us an excellent framework
to focus the discussion during the workshop. A plan based on stories
tells us exactly what we need to focus on for the next iteration and
which benefits we need to fulfill. Having this clearly defined helps
focus the specification workshop on particular customer expectations,
giving us a platform to discuss which examples are really important

170

User stories and acceptance tests

for this phase of development and which should be left for later. This
helps a lot to eliminate just-in-case code.
In essence, user stories are the scope for the project, facilitating longterm planning and helping us get the big picture of what needs to be
done. Acceptance tests are the detailed specifications that are ironed
out only before the implementation of a particular story or a set of
related stories. With regard to three Cs of user stories, the specification workshop is the conversation and acceptance tests coming out
of the workshop are the confirmation.

Stuff to remember
Agile acceptance testing and user stories complement
each other incredibly well.
You do not have to plan projects with user stories, but
there are significant benefits if you intend to introduce
agile acceptance testing.
User stories focus on customer benefits, so it is easier
for customers to plan based on them.
Group stories into deliverables by business goals and
make sure that each deliverable can go to production.
User stories are the scope. Acceptance tests are the
specification.

171

172

Chapter 10

Tools of today
Agile acceptance testing relies on test automation, so much so that
the tools that we use for automation dictate the form in which tests
are recorded and guide us in development. In this chapter, I describe
several tools that you can use for test automation while still keeping
tests in a form that can be understood by both business people and
software implementation teams.
Tools and technologies come and go, but ideas and practices stay with
us. In fact, the incompleteness of today's tools is one of the major
obstacles to the adoption of agile acceptance testing. I expect tools to
improve considerably in the near future, so giving you detailed
instructions on how to use them in this book would not be effective.
You will be able to find more complete and more up-to-date user
manuals for tools online.
Instead, I want to present an overview of some of the most popular
or interesting tools offered today, as a short tourist guide that will
inspire you to explore further and tell you where to look for more
information. This chapter is a bit technical, so feel free to skip parts
of it if you are only interested in the business side of the story.

FIT
Framework for Integrated Testing (FIT) was the first popular acceptance testing framework, originally developed for Java by Ward
Cunningham in 2002. At that time, Ward Cunningham called it a
tool for enhancing collaboration in software development.1 One of
the central ideas of FIT was to promote collaboration and allow
customers, testers and programmers to write and verify acceptance
tests together.
1

http://fit.c2.com/

173

Tools of today

FIT works with a tabular model for describing tests. A typical FIT
test is shown in Figure 10.1. Test inputs and expected results are
specified in a table, with expected outcomes ending with a question
mark in the column headings. This tabular form makes it very easy
to write and review tests. Results are presented in the same format.
If a particular expected value does not match the actual outcome, FIT
marks the cell red and prints out both the expected and actual value.
FIT simply ignores everything outside of the tables, so additional
documentation, project notes, diagrams and explanations can be
easily bundled with acceptance tests, providing deeper insight into
the problem domain and helping people understand and verify test
results. All this helps to evolve tests with the code more easily.

Figure 10.1. A typical FIT test table

These tabular tests are excellent for calculation rules, but they are not
really good for describing workflows or stories. Rick Mugridge has
developed an extension to FIT called FitLibrary that allows us to
specify flow tests as stories in a format similar to English prose (see
Figure 10.2). Although it is technically a separate extension, FitLibrary
is now considered by most people to be part of the standard set of test
types. FIT and FitLibrary together provide about ten different ways
to specify tests with tables, including lists, sets, stories and calculations
based on column and row headings. This makes FIT very flexible and

174

FIT

because of this it remains the most popular tool for acceptance testing
despite much recent competition.

Figure 10.2. FitLibrary flow test

FIT tables connect to the domain code using a very thin glue code
layer. These layers are called fixtures. A fixture is effectively more an
integration API then a testing API. The fixture tells FIT how to
interpret a particular table, where to assign input arguments, how to
execute a test and where to read actual test outputs. The example in
Figure 10.3 is a Column Fixture, which maps table columns to public
fields, properties and methods of the fixture class. In general, FIT
requires very little extra code for testing just enough to provide a
sanity check for the underlying interfaces. Often, FIT fixtures
constitute the first client of our code.

175

Tools of today

Figure 10.3. FIT fixtures are the glue between tables and
domain code

FitNesse
FIT makes it easy to run tests, but does not provide a way to create
or manage them. The original idea was to write tests in Word, Excel,
or any tool that can output HTML. FitNesse (see Figure 10.4) is a web
wiki front-end to FIT developed by Robert C. Martin and Micah
Martin of Object Mentor in 2005. Today it is the most popular choice
for running FIT tests. It provides an integrated environment in which
we can write and execute tests and speeds up the job with quite a few
useful shortcuts. Although FitNesse is also written in Java, it is not
heavily integrated with FIT, but executes it as an external program.
This is very useful, as it enables us to plug in different test runners.
After the FIT-FitNesse combination became popular in the Java world,
test runners were written for other environments including C++,
Python, .NET and SmallTalk. FitNesse is a web-based server, allowing
easy collaboration. Business analysts and other non-technical people
do not have to set up any software in order to use FitNesse. Any
browser will do just fine. FitNesse also allows us to organise tests into

176

FitNesse

test suites, reuse a shared set-up or tear-down for the entire test suite
and include test components, making tests easier to manage and
maintain long-term.

Figure 10.4. FitNesse is a web wiki front-end to FIT

However, it is not without its problems. FitNesse was designed as a

stand-alone network server with its own version control system, but
most people simply did not want to use it this way and continued to
store tests in the main project version control system. Although the
internal version control system in FitNesse can be turned off, because
of the way that tests are stored on disk some of its features don't really
work with folder-based version control systems such as Subversion
and CVS. The way that tests are organised in FitNesse at the moment
also does not scale well when you need to deal with thousands of tests
in deep hierarchies.
In general, FIT and FitNesse have a relatively short learning curve
and allow people to get started with acceptance testing really quickly.
The tabular form of specifying tests is very concise and I personally
prefer it to free-form text or keyword-based systems, although it is
arguably not the best to describe workflows.

177

Tools of today

Alternative tools
A number of alternative test runners and test management tools for
FIT have emerged in the last several years. FitClipse by DoubTech2
is an Eclipse plug-in that manages FitNesse tests. BandXI also offer
an Eclipse plug-in3 that enables developers to start and stop FitNesse
from Eclipse and execute tests. Jeremy D. Miller has written
StoryTeller,4 which provides a way to run and manage tests in Visual
Studio. Rick Mugridge is working on ZiBreve,5 a Java IDE for story
tests with WYSIWYG editing and support for refactoring tests. Jay
Flowers has written a plug-in6 for TestDriven.NET that runs FitNesse
tests from Visual Studio.
FIT, FitNesse and most of the supporting tools are open source. For
people requiring commercial support and integration with commercial
tools, there is Green Pepper7 by Pyxis Software. Green Pepper is built
on similar concepts to FIT and FitNesse, but it integrates nicely with
Confluence and JIRA from Atlassian, providing a more integrated
way of managing tests and relating them to development tasks and
issues. An interesting feature of Green Pepper is that it supports
multiple versions of a single test, so that you can mark which version
is actually implemented and which version is being implemented for
the next release, allowing you to manage more easily the transition
from acceptance tests to regression tests and back.

More information
For more information on FIT and FitNesse, see FIT for developing
software[9] by Rick Mugridge and Ward Cunningham and my book
Test Driven .NET Development with FitNesse[22]. The Fixture Gallery8
2

http://www.doubtech.com/development/software/projects/?project=1
http://www.bandxi.com/fitnesse/index.html
4
http://storyteller.tigris.org/
5
http://www.zibreve.com/
6
http://jayflowers.com/WordPress/?p=157
7
http://www.greenpeppersoftware.com/en/products/
8
http://sourceforge.net/projects/fixturegallery
3

178

Concordion

is a free PDF guide that explains all the most popular test (fixture)
types, offering advice on when to use them, when not to use them and
how to save time and effort when writing tests. For more resources,
see the main FitNesse web site http://www.fitnesse.org and the
community-maintained public wiki http://www.fitnesse.info.

Concordion
Concordion, developed by David Peterson and released under the
Apache open source license, is an interesting alternative to FIT. Like
FIT, Concordion uses HTML documents as executable specifications
and requires some glue code (also called fixture code) to connect the
executable elements of the specification to the domain code. Unlike
FIT, Concordion does not require the specification to be in any
particular format you can write examples as normal sentences,
without any restrictions.
Concordion is really simple. Its instrumentation only allows
programmers to set global test variables, execute fixture methods and
compare actual results with expected values. Programmers can use
special HTML element attributes to mark words or phrases that are
used as test inputs or compared to test results. Web browsers just
ignore unknown element attributes, so Concordion test instrumentation is effectively invisible to people that are not interested in test
automation. For repetitive specifications and calculation rules,
Concordion also supports attributes for tables similar to those of the
FIT Column Fixture.
Currently, Concordion only supports Java fixtures, since it actually
works as a JUnit extension. This provides direct integration with
JUnit test runners, making it much easier to execute Concordion tests
from popular development environments and integrate them into
continuous build systems. It also shortens the learning curve required
to start using Concordion. The fixture is simply a JUnit test class,
with the same name as the HTML file containing the test.

179

Tools of today

Concordion's fixture model does not depend on inheritance so it is

easier to learn than FIT and more flexible. On the other hand,
Concordion lacks the extensibility of FIT and its powerful type
adapters and cell handlers that enable it to bind domain objects and
business services directly to the specification.
When a test is executed, Concordion runs through the files and
executes commands, comparing expected outputs with actual results.
The JUnit test run tells you whether all tests passed or there were
failures. Concordion also saves the results to disk in HTML, making
it easier to see what actually went wrong. You can see an example of
a test result on Figure 10.5.
Concordion does not have a test management tool and relies
completely on the programmer's IDE to manipulate and execute tests.
I have mixed feelings about this. Although file management within
an IDE makes it much easier to put tests in the same version control
system as the domain code, I miss the ability to use common set-ups
and test components with macro variables such as are available in
FitNesse. It also puts tests completely under the control of programmers.
Concordion takes some ideas that have evolved as best practices in
the FIT-FitNesse arena and makes them very explicit, actively
discouraging other ways of working. For example, acceptance tests
have to be stored in the same folder structure as the code, making it
easy to include them in the same version control system. There are
no prebuilt test building blocks that would encourage people to reuse them, often leading to large unreadable test scripts (see the section
Distilling the specifications on page 78).
What worries me slightly is that developers need to add non-standard
HTML attributes to the test page. This smells to me of a hand-over
of tests from business people to developers at some point, which is a
practice that I don't approve of at all. In my opinion, tests should be
shared between the whole team and not handed over. I am sure that
the problem can be avoided with some discipline, and there are
probably visual HTML editors which business people can use that

180

JBehave

will not ruin hidden test instrumentation, but I would still like to see
a proper test management tool for business people and testers.
Exposing tests in an IDE is great for developers, as they can debug
and troubleshoot tests easily, but testers and business people need to
view tests at a higher level.

Figure 10.5. Concordion works with any text in a HTML

document

Concordion has a really good web site with a helpful tutorial and lots
of examples. For more information on this tool, and to download
links and examples, see http://concordion.org.

JBehave
JBehave is the original behaviour-driven development tool, written
for Java by Dan North, Liz Keogh, Mauro Talevi and Shane Duan.
Behaviour-driven development uses a specific format of test scripts
specified in the given-when-then format (see the section Working
with business workflows on page 57). Here is a typical JBehave test
script, taken from the JBehave source code repository:

181

Tools of today
Given a stock of prices 0.5,1.0
When the stock is traded at 2.0
Then the alert status should be OFF
When the stock is traded at 5.0
Then the alert status should be OFF
When the stock is traded at 11.0
Then the alert status should be ON

In a similar way to Concordion, JBehave plugs itself into the JUnit

unit testing tool and works on a text file with the same name as the
fixture class. In the JBehave jargon, both the series of given-whenthen steps and the related glue code between test scripts and domain
code are called a scenario. Individual elements of the given-when-then
workflow are called steps. Each step is implemented as a Java method
and annotations are used to match lines of text from the test script
to the methods and their arguments.
JBehave test scripts are in plain English, making them really easy to
understand. There is no hidden test instrumentation, which means
that test scripts are easy to maintain and that business people can use
any text editor or word processor to maintain the scripts. This also
means that there is no way to specify lots of related specifications in
a compact format such as a FIT table. I find that the given-when-then
format is better for describing workflows and unique cases, whereas
FIT tables are better for describing calculations and state-machine
changes with lots of related rules.
Much of what I've written about Concordion applies to JBehave as
well. Since JUnit is used under the hood, JBehave has a shorter
learning curve technically and it integrates instantly with almost every
IDE, build tool and continuous integration system. Test scripts are
kept in the same folder structure and source code control system as
the production code. There is no specific test management tool and
test scripts are primarily under the control of the programmer's IDE.
JBehave feels much more like JUnit than any other tool described
here, so if you are a programmer and want to start diving into
acceptance testing, it might be the easiest one to learn. Compared to
JUnit, it has test scripts externalised into text files and changes the

182

TextTest

language from tests and assertions to what the domain code should
do, emphasising the fact that a test in this case is a specification. Unlike
Concordion, the test script instrumentation is in the fixture class.
This means that test scripts are unaffected by code refactoring, but
it also means that you have to write a lot more code than with
Concordion. Unlike Concordion, there is no standard way to specify
lots of related examples with table attributes. You can, however, extend
the way JBehave parses scenarios by implementing your own converter
to handle such cases. Currently JBehave developers are working on
a standard extension that would allow you to specify related examples
as a table under the other examples include heading.9
You can download JBehave from http://jbehave.org. If you are not
using Java, there are similar tools for different platforms and
programming languages. In fact, the Ruby version, RSpec, seems to
be a lot more popular among Ruby programmers than JBehave is in
the Java community. At the Agile 2008 conference a new Ruby BDD
tool called Cucumber was announced, with two very interesting
features. The first is support for internationalisation, making it easier
to apply BDD in another language besides English. The second
interesting innovation is the ability to supply additional combinations
of inputs and expected outputs for a scenario using a simple table
syntax appended to the classic RSpec scenario template, very similar
to a classic FIT column fixture. This could make writing and maintaining BDD tests much easier as you would no longer have to copy
and paste scenario descriptions to use different arguments.

TextTest
TextTest is a tool for managing system behaviour changes written
by Geoff Bache, and it stands out from all the other tools mentioned
here because of its way of working and its unique feature set. Geoff
Bache explains10 it as a move away from trying to write a sequence
of assertions that entirely describes the correct behaviour of the
program:
9

http://www.testingreflections.com/node/view/7339
in a private e-mail

10

183

Tools of today

Systems change over time and so does their behaviour: why not
have a testing tool that assumes this as normal? This view seems
to me entirely in line with an agile view of software
development, though it is a bit of paradigm shift if you're used
to testing with assertions and APIs [...] I don't see this as
different or mutually exclusive to testing in development (if a
program isn't in development, there isn't so much point testing
it). My attitude is that however incomplete the program, it does
something today that is better than nothing: so we manage the
changes in its behaviour and make sure that we always move
forwards, never backwards, when we change the code.
Unlike FIT and Concordion, which inspect the system using an
integration fixture API, TextTest inspects the system by analysing
log files and other outputs generated during test runs and and
comparing them to expected results. Because of this, TextTest is much
better for verifying business workflows and integrations. FIT and
Concordion can only confirm the results of a function call through
the external API. TextTest can confirm that the correct process was
executed during the function call, by analysing individual steps and
verifying the flow of information.
TextTest also supports multiple active versions of the same test and
integrates with xUseCase tools for recording and replaying user
interface tests. xUseCase libraries differ from standard UI recorders
because they associate recorded actions with use case command names
rather than captured UI system events. Use case commands are named
in domain terms, as suggested by business people, testers and
developers. Test scripts can then be written using these domainspecific names, not by capturing system events. This allows us to focus
user interface test scripts on what the system is intended to do in
terms of the domain, without getting involved in implementation
details. The result is that xUseCase scripts are much easier to understand and manipulate and are much closer to the domain. This is
similar to the domain-specific testing language approach of Mickey
Phoenix and SolutionsIQ, explained in the section Don't describe
business rules through the user interface on page 91. TextTest integ-

184

TextTest

rates with PyUseCase11 and JUseCase12 providing this user-interface

recording functionality to Python and Java programs.
Although TextTest is intended to be primarily a regression test tool,
it allows us to define future business workflows as sentences in a text
file and then use them as acceptance tests for development. In the
context of the topic of this book, the most interesting features of
TextTest are its ability to analyse business workflows and its domainoriented user interface recording and replay functions.
Since it works on text files, TextTest can inspect software written in
any language on any platform, so if you are working in a legacy
environment or with a non-standard platform, TextTest might be
your only choice for acceptance testing. It could also prove useful as
a tool for retrofitting regression tests into a legacy system, because it
is much less intrusive than tools that depend on an external API.
In addition to functional testing, TextTest has some unique features
such as built-in performance test support and the ability to record
and replay messages to other programs or systems to help mock out
external dependencies for testing.
TextTest has a console interface for automation and a graphical
interface for easier use (see Figure 10.6). The graphical interface also
serves as a test management tool, allowing you to group tests into
hierarchies. TextTest is itself written in Python, so it can be extended
by writing Python modules that preprocess or postprocess inputs and
outputs to customise the test execution for your particular application.
To get the most out of this approach to testing, the key is to create
logs with an appropriate level of detail. Geoff Bache recommends that
good logs should be:
as readable as possible for someone who knows the system but not
the code

11
12

http://sourceforge.net/projects/pyusecase/
http://jusecase.sourceforge.net/

185

Tools of today

detailed enough to describe the important things that the system

does
not so detailed that they become hard to read and excessively
volatile
This ultimately leads to much more focus on designing the log files
and having logs that are understandable to non-programmers. I find
this side-effect very interesting, as it should make log files much more
usable for support and troubleshooting in production as well.
For more information on TextTest, see http://www.texttest.org/. I
especially recommend the Publications page, which has slides from
Geoff Bache's presentation from the Agile 2008 conference and the
Minesweeper Kata screencast.

Figure 10.6. TextTest graphical user interface

186

Selenium

Selenium
Selenium is a web browser automation tool that can manipulate and
validate HTML document object model elements. The three most
common uses of Selenium are automating web user interface tests,
verifying web site compatibility with various browsers and platforms
and periodically checking whether a site is online and works correctly.
In the context of the topic of this book, Selenium is interesting as an
engine that drives acceptance tests through the user interface.
The Selenium project started as a JavaScript functional test runner
in 2004 created by Jason Huggins, Paul Gross and Jie Tina Wang,
who were working for ThoughtWorks at the time. It is an open source
product, actively maintained by ThoughtWorks, with a large
community of supporters. Since 2004, it has significantly evolved in
terms of functionality and tool support. In fact, additional tool
support and integration options are two of the best features of
Selenium.
The core of Selenium is written in JavaScript and HTML and it is
compatible with all major browsers, including Internet Explorer and
Firefox. It runs on Windows, Linux and Mac. Selenium works by
embedding control scripts into a browser window frame and automating another window frame with the scripts. An example Selenium
screen is shown in Figure 10.7. The top half of the screen is the
Selenium control interface: the top-left corner shows the current test
suite, the middle part shows the current test and the top-right corner
contains the panel that controls tests. The bottom part of the screen
shows the web site under test.
Selenium test scripts are written as tables, with table rows representing
test steps. The first cell in a row contains a command keyword, the
second and third cells contain command arguments. This test language
is known as Selenese. For an example that submits a search form and
validates the output, see Figure 10.8. Selenese has many different
keywords to simulate clicks, type text into fields, load pages, evaluate
text in elements, check for alerts and manipulate and inspect HTML

187

Tools of today

element properties.13 Selenium can identify document object model

(DOM) elements based on their IDs, names and XPath expressions,
so its test language is very flexible in terms of working with HTML
pages. You don't have to be a programmer to write Selenium scripts,
but you do need technical knowledge about the DOM and Selenese
is very low-level and technical. Therein lies one of the biggest problems
with Selenium. Writing and maintaining Selenium scripts by hand
is probably not the best idea, even if you are a programmer, because
small changes in the user interface can break many tests. Because of
this, Selenium is rarely used on its own. It is mostly combined with
other test management tools or wrapped into interfaces that make
test scripts easier to use and maintain.

Figure 10.7. Selenium test runner

Selenium has a visual test automation tool that supports record-andreplay operations, called Selenium IDE (see Figure 10.9). The tool is
aimed at less technical people who do not want to write Selenese
scripts by hand and works as a Firefox plug-in (Internet explorer is
not supported currently). The Selenium IDE comes with good

13

see http://selenium-core.openqa.org/reference.html for more information

188

Selenium

command reference documentation, point-and-click helpers and the

ability to save and load tests from the disk.

Figure 10.8. A typical Selenium test

For more flexible automation, there is Selenium Remote Control.

Remote Control is an add-on that provides the glue between the core
browser automation engine and .NET, Java or Python code. It works
as a Java server process that accepts commands over TCP and launches
a browser to execute them. Remote Control enables us to write and
run Selenium tests from almost all popular test frameworks,
combining them with integration test functions that prepare, clean
up or validate domain objects. It also allows us to remotely execute
Selenium tests on dedicated machines, making it possible to automatically run the same test in multiple browsers on different operating
systems.
The biggest problem of Selenium Remote Control is speed of execution
the interface between the Remote Control server and JavaScript
running in a browser is a real performance hog and makes it very
ineffective to run Selenium-based tests as part of the basic build
process. A possible solution for this problem in the Java world is to
use TestNG with parallel test execution. The Selenium Grid project
is another possible solution. It automatically distributes Selenium
tests across multiple machines and comes with built-in support for
Amazon's EC2 Cloud.
Download Selenium from http://www.openqa.org. There you will
also find find code examples, a command reference and installation
and usage instructions for the core library, Remote Control and
Selenium Grid.

189

Tools of today

Figure 10.9. The Selenium IDE allows testers to record

and replay tests

Tools based on Selenium

CubicTest is an interesting tool that allows us to write and manage
Selenium tests as visual page transitions (see Figure 10.10). It also
offers the ability to generate mock web sites based on test scripts, so
it might be an interesting tool for prototyping user interface workflows. CubicTest is an open source plug-in for the popular Eclipse
IDE. For more information and to download, see
http://www.cubictest.com.

190

Tools based on Selenium

StoryTestIQ, a wiki-based Selenium test management system by

SolutionsIQ, is a mixture of Selenium and FitNesse. It offers easier
test management and the ability to run SQL commands to prepare
the database for testing through the user interface. It integrates with
continuous integration tools. You can download StoryTestIQ from
http://storytestiq.solutionsiq.com.

Figure 10.10. CubicTest manages tests as visual page

transitions

WebTest fixtures are a library of additional FIT fixtures for Java and
.NET which I wrote to allow business analysts and testers to specify
Selenium tests without actually learning the Selenese language. The
fixtures convert English sentences such as user clicks on Cancel and
Page contains text Hello Mike into Selenium Remote Control
commands, enabling non-programmers to write and maintain user

191

Tools of today

interface tests from FitNesse. For more information, see

http://fitnesse.info/webtest.

Stuff to remember
There are many different acceptance testing tools today
on the market.
Start by evaluating FITFitNesse, Concordion and JBehave
for business domain testing.
Selenium, CubicTest and StoryTestIQ might help you with
UI tests.
To maintain UI tests more easily, use page objects or
domain-specific keywords to describe actions in a business language.
TextTest can be an interesting choice for inspecting
workflows and building up regression test suites for
existing systems.

192

Chapter 11

Tools of tomorrow
Tool support for agile acceptance testing is today good enough for
us to get our work done, but not good enough for our work to be
really efficient. Early adopters of the practice were forgiving and ready
to overlook problems with tools, working around them or implementing their own solutions. However for this practice to become widespread, we need better tools that give us a shorter learning curve and
allow us to do our jobs more efficiently. As the practice evolves, the
tools need to evolve too. Here are some ideas that I would really like
to see in our tools in the near future.

Domain-specific languages
The idea of domain-specific languages (DSL) seems to be gaining a
lot of momentum currently. The basic idea is to create a specific miniprogramming language, often based on some action-oriented
keywords, focused on a particular business domain and even tailored
to the needs of a particular customer. We then use this mini-language
to describe business processes and constraints. Business people and
domain experts can then effectively participate in programming
even if they don't want to write in the language themselves, they will
at least be able to read and understand it. So domain-specific
languages can improve communication in a project. Another rationale
behind this idea is that domain-specific language constructs allow us
to describe the business problem more efficiently than with generalpurpose languages. The mini-language is ultimately implemented in
a general purpose language such as Java or C#, so that process definitions written in the business language can be integrated with the
infrastructure.
This idea is very close to the concept of writing acceptance tests so
that business people can understand them. In agile acceptance testing,
the acceptance tests are written in business language and serve as a
193

Tools of tomorrow

specification for the code. Verifications and actions related to the

business language keywords are then implemented in a generalpurpose language, for example Java or C#.
A domain-specific language does not have to be based on keywords
in text. FIT, Concordion and similar tools actually already give us a
way to use domain-specific languages without calling them by this
name. FIT, for example, enables us to use tables to create a DSL. The
automation part (the fixtures) can be viewed as an implementation
of the domain-specific language described by the constructs used in
acceptance tests.
With some lower-level tools such as user interface recorders, tests
are described as series of small technical steps, making it hard to see
the wood for the trees. Some companies such as SolutionsIQ are
pioneering the idea of domain-specific testing languages applying
the idea of domain-specific languages to encapsulate technical
complexity into something that business people can read and write.
The constructs of such languages are used to describe user-interface
acceptance tests in a form readable by business people. The implementation of the constructs encapsulates all the technical details,
hiding them in the code. Instead of a web page test script with thirty
steps to populate and submit a registration form, confirm terms and
conditions and verify that the registered e-mail address is correct, we
might just use the register-customer keyword in the test specification.
The benefits of this approach are better readability and easier
management and maintenance of acceptance tests. Developers can
use domain-specific keywords to discuss tests with business people
and user interface changes will be localised to individual keyword
implementations, they will not propagate to domain tests. As userinterface testing tools evolve and speed up, this idea might prove to
be more and more interesting. Some automation tools such as
JUseCase and PyUseCase provide the infrastructure for this for thickclient applications.
As agile acceptance testing gains in popularity this concept will surely
evolve. Domain-specific languages and their counterparts in testing
are aleady becoming more and more popular witness the support

194

Different user interfaces for different roles

for dynamic languages in Microsoft's .NET CLR and the increasing

number of Java JRE-based languages.
Adam Geras held a talk during the second Agile Alliance Functional
Testing Tools workshop, during which he suggested that instead of
a testing tool or framework a good future base for acceptance testing
using domain-specific languages would be a new virtual machine.1
The virtual machine would provide support for different testing
contexts, including test-first, test-after and exploratory testing. Such
a virtual machine would give us full flexibility to write tests and
specifications in domain-specific languages, but would not impose
any specific functional form or limitations.

Different user interfaces for different

roles
Current test management tools try to provide a compromise between
technical efficiency and user-friendliness in order to be usable by
developers, testers and business people. Jennitta Andrea has written
an article called Envisioning the Next Generation of Functional Testing
Tools2 in which she suggests that future tools should provide better,
more focused interfaces for different people depending on their roles.
Jennitta writes that a functional test has to be easier to read and locate
than code. In addition, as a single test often describes just a part of a
business rule, related tests have to be easy to locate and review.
Different people expect different things from a test. Business people
and domain experts are mostly concerned with verifying the correctness of the specification in the test. Testers are mostly concerned with
running the test and getting meaningful execution results. They also
want an efficient way to identify and describe bugs. Developers are
concerned with understanding the test as a specification for development or bug-fixing. Business analysts and developers are often
1

http://video.google.com/videoplay?docid=-6298610650460170080&hl=en
See [23]. Also see http://www.jennittaandrea.com/wp-content/uploads/2007/04/envisioningthenextgenerationoffunctionaltestingtools_ieeesw_may2007.pdf.
2

195

Tools of tomorrow

concerned just with the current state of tests, not so much about the
previous test runs. Testers are typically interested in the results of
earlier test executions and previous versions of tests as well.
I'm not so sure that all these roles can be reconciled in a single user
interface, and I'd like to see multiple user interfaces that focus on
providing the best individual benefits for particular roles. For example,
I would expect better integration with traditional integrated development environments for programmers, so that they can debug and
troubleshoot code under test while they are implementing parts of a
specification or bug-fixing. Testers, on the other hand, should not
have to run tests from a programmer IDE, but from an interface that
gives them a quicker global overview with fewer details. Business
people should be able to write and review tests in a visual WYSIWYG
environment similar to a word processor or spreadsheet. I would like
future tools not to attempt to be all things to all men, but to provide
all the features that a particular user needs to do his job efficiently,
while allowing others to use different tools.
Even though some IDEs such as Eclipse offer a rich client application
development framework, I still think that applications based on IDE
frameworks are too technical for business people to use, and that we
need a proper business view of the tests. After all, this is why FitNesse
is still the most popular test management system for FIT tests in spite
of all its problems.

Propagating the effects of changes

Refactoring is the practice of improving the design of existing code
without changing its functionality, today regarded as one of the key
agile development practices. If you are not a technical person, consider
it as periodic house-cleaning after development. The comparison is
not entirely correct because refactoring is typically a continuous
process and not a once-off effort, but it is good enough for our
purposes. Most integrated development environments today support
refactoring automation, so that changes are automatically propagated
to related files. For example, when part of an object is extracted into

196

Propagating the effects of changes

a new object, or when a function name is changed, the change is

automatically implemented in all the places where the object or the
function are used. This automation makes such changes much cheaper
and less error prone.
Unfortunately, since acceptance test descriptions are not kept in the
code but in external files, such refactoring does not affect acceptance
tests. Some tools decouple the test description from the code using
test fixtures. Automated tools propagate changes applied in the
domain code to the fixture code as well, while leaving the fixture
interface unchanged. This allows tests to still pass after refactoring,
but the language in the tests is no longer consistent with the language
in the domain. This makes tests harder to understand and goes against
the principle of using a consistent ubiquitous language on the project.
With tools that skip the fixture part and hook test scripts directly to
domain objects or analyse the raw unprocessed output of domain
classes, refactoring domain code breaks the tests.
In both cases, tests need to be updated to be kept consistent with the
domain code. Regardless of the tool, acceptance tests become out-ofdate after code refactoring. Currently they need to be fixed manually.
This is one of the major pain areas of acceptance testing today.
ThoughtWorks were working on a tool called Profit3 that converted
FitNesse tests to Java code, which would be affected by refactoring,
but this project seems to have been abandoned. The last update was
more than two years ago and it was still marked as in planning.
However, this is an interesting idea, although I suspect that turning
tests into code will make them much more technical and less understandable. While I was writing this book, ThoughtWorks have
announced plans to release a new tool called Twist4 that promises to
bring better refactoring support.
I expect that future tools will have much better support for refactoring,
plugging into the integrated development environments and automat3
4

http://sourceforge.net/projects/profit
http://studios.thoughtworks.com/twist-agile-test-automation/

197

Tools of tomorrow

ically propagating changes to acceptance tests. This could work both

ways, with business people updating tests and these updates
propagating to the domain code.

Direct domain mapping

Most acceptance testing tools today allow us to describe tests in a
human-understandable language and then have a translation layer
that converts it into domain API calls. Acceptance tests are a really
good way to create and evolve an ubiquitous language for the project,
which helps a great deal to improve communication between business
people and technical people on the team (see the section Building a
domain language on page 66 and the section Evolve the language
consistently on page 134). Using this approach, both the domain code
and acceptance tests use the same names and constructions, so the
translation API becomes just boiler-plate code, but it is still code that
has to be written and maintained.
Both FIT.NET and Java FitLibrary have started to move towards using
a rich domain model that is more effective than the traditional
mapping of each step into a fixture or a DoFixture method. Domain
fixtures and domain adapters in FIT.NET and Java FitLibrary utilise
rich domain objects without the need to write boiler-plate fixtures.
These mappings are still in the early stages of development and are
not comprehensive adapters still have to be written for specific
domains. During the Agile 2008 conference, Mike Stockdale organised
a mini-session where he and Rick Mugridge presented some new
features and ideas that they were working on at the time. The session
led to a very interesting discussion on whether we could produce types
of domain adapters and domain fixtures that would allow FIT to
connect directly to most domain services and objects without the need
for any fixtures.
Moving forward in this direction, this approach would allow acceptance test automation tools to directly utilise domain-driven design
concepts such as repositories, business services, entities and value
objects. This would remove the need to implement any generic fixtures

198

Better editors

to connect test scripts to domain code. Only genuine test-specific

code would need to be developed and maintained. By eliminating the
translation layer, this approach also promotes the use of the ubiquitous language.
A major problem with connecting test scripts directly to domain code
at the moment is that it makes tests harder to maintain, because test
scripts are not isolated from changes in the code. On the other hand,
test scripts and domain code have to be kept synchronised anyway to
make the most of the ubiquitous project language approach, even if
they are isolated. Better support for IDE integration and refactoring
will make test scripts easier to maintain, allowing us to obtain in full
the benefits of directly connecting to the domain. For some ideas that
came out of the workshop during Agile 2008, see my article FIT
without Fixtures.5

Better editors
During the second Agile Alliance Functional Testing Tools workshop,
Elisabeth Hendrickson called for better editor tools rather then just
support for refactoring.6 Her idea was that future tools must reduce
the cost of changing tests, allowing us to efficiently reconcile new
expectations with those already expressed as acceptance tests. Better
editors will reduce the cost of change and facilitate change in tests
rather then impede it. Elisabeth Hendrickson pointed out two key
features that future editors should have:
They should allow people to quickly identify mismatches between
expectations in their minds and those written down as acceptance
tests.
They should make it easy to change tests to reconcile mismatches
easily and quickly.
With better editors like these, the lack of support for carrying through
refactoring into acceptance tests would not be such a problem, because
5
6

http://gojko.net/2008/08/12/fit-without-fixtures/
http://video.google.com/videoplay?docid=8565239121902737883&hl=en

199

Tools of tomorrow

we could easily implement changes in acceptance tests to reflect

domain changes after refactoring.
Jennitta Andrea has suggested[23] that better test editors should
provide intelligent language support that works with domain-specific
testing languages, in the same way as intelligent code helpers work
in developer tools.

Better test organisation

FitNesse has an internal version control system that has turned out
to be more of a stumbling-block than a useful feature. Most people
want to keep tests and domain source code in the same repository
under the management of a single version control system. Other tools
don't even try to tackle this issue and rely on the IDE to handle version
management. A key feature of future tools, especially if they offer
multiple views of the same tests, has to be proper version control
integration with popular source-code control systems. A FitNesse
plug-in intended to allow this has been in development for a few years,
but as of now it is still not complete. This feature is not so important
for programmers, because most IDEs already support many version
control systems. But it will be a key feature of business-oriented and
tester-oriented interfaces to acceptance testss.
In addition to this, future tools need to support multiple active
versions of a single test. A single version of the test for each branch
of source code is not enough. With acceptance tests that facilitate a
change to a system, we might use several versions of the same test
with the same source code branch. One version would describe what
is currently implemented and be used for regression testing. A second
version would describe where we want to be once the change is
implemented and be used for guiding development. At the moment,
having two versions of a single test often requires copying the test
into a development space and then having two copies of the same file,
with testers and the continuous integration environment focusing
on the old file and developers working with the new file. Green Pepper
and TextTest have direct support for multiple test versions, so that

200

Visual workflow descriptions

we can mark a revision as implemented or in development and then

just execute tests marked with a particular version. I would like to see
this feature in other test management tools as well, because it would
make test management much easier.
Rick Mugridge has started work to implement test tags for SuiteFixture in FitLibrary. Test tags are arbitrary categories that can be
assigned to tests so that we can list, execute and group tests, providing
a more flexible mechanism to manage tests then the traditional filesystem-based hierarchy. I would really like to see this implemented
in test management tools such as FitNesse rather then in a libraryspecific test runner like FitLibrary. Test tags woule meet Jennitta
Andrea's requirement that related tests should be easy to locate.

Visual workflow descriptions

Workflows are, at the moment, hard to describe and check properly
with current acceptance testing tools. Most people are used to
discussing and designing workflows with visual diagrams, but today's
tools require us to describe workflows as sequences of steps in text.
This makes it much harder to understand even simple straightforward
workflows. Flows with lots of branches and conditionals, described
as text steps, truly exemplify the proverb that a picture is worth a
thousand words. Pictures are definitely a more effective way to design
and communicate workflows. However, currently they are not
executable.
One of the areas where acceptance testing tools will probably evolve
in the future is visual workflow exploration and specification. Some
tools for this already exist. CubicTest, described in the section Tools
based on Selenium on page 190, has a visual page transition model to
specify web workflows. Brian Marick has written a library that reads
OmniGraffle graph documents and converts them into executable
Ruby on Rails tests.7

http://www.exampler.com/blog/2007/07/13/graphical-workflow-tests-for-rails/

201

202

Part IV. Effects on all of us

In this part, I try to answer the most common questions I get from business analysts, testers and developers on how agile acceptance testing will
affect their roles. I also dispel common misconceptions and address fears
and issues that people often have about agile acceptance testing because
they have been misled by partial information.

Chapter 12

Effects on business
analysts
During the QCon 2007 conference in London, Martin Fowler and
Dan North organised a session called The Crevasse of Doom,1 to
discuss the communication gap between business users and software
developers. Martin Fowler called this gap the biggest difficulty we
face in software development today. He divided the strategies for
crossing the gap into two groups: the ferry and the bridge.
The ferry is the traditional way to think about a communication
problem. Fowler said that developers are often perceived to have poor
communication skills, dress in T-shirts and talk in techno-jargon, so
business people who dress in smart suits and talk in their business
jargon think that they could not possibly communicate with
developers. Professional intermediaries are hired as translators, who
ferry information between the two groups. The problem with this
approach is, as Fowler put it, that it is not a high-bandwidth form of
communication and it makes no difference to the tradional gulf
between developers and business people.
The bridge approach provides a structure that allows people to cross
the gap directly, facilitating direct communication between the two
groups. Instead of having professional translators, we work on
fostering communication and building links between the two worlds,
making a single world out of them. The bridge approach provides a
much higher bandwidth and less distortion. It also offers an opportunity for people with different skills to work together to find better
ways of solving business problems.

http://www.infoq.com/presentations/Fowler-North-Crevasse-of-Doom

205

Effects on business analysts

In the traditional requirements gathering process, a business analyst

is a conduit of information. He or she is an expert hired by management to transfer and translate information because it is believed that
developers and business people cannot understand each other. I think
that this assumption is essentially flawed and I hope that, if you have
read the preceding chapters, I have also convinced you to believe the
same. This by no means makes the business analyst obsolete, but it
affects the role that an analyst plays in the process. As Fowler put it,
the traditional role of a business analyst is to be the ferry man. With
agile acceptance testing, analysts become bridge builders. They are
no longer the conveyor of information, but facilitators that enable
people to share knowledge.
Rather than producing and handing over documents, the role of a
business analyst is to organise the transfer of knowledge and chase
open issues. The business analyst is a great choice for the facilitator
of specification workshops. Very often, a business analyst is charged
with owning and maintaining acceptance tests after workshops as
well.

Benefits for business analysts

In the introduction, I listed six significant benefits of agile acceptance
testing for business analysts. As a quick recap of Part II, let's look at
these benefits again and see how the practices described deliver them.

Developers will actually read

specifications
Ellen Gottesdiener wrote that collaborative requirements workshops
commit participants to the workshop results more and promote their
sense of ownership of the deliverables and the system ([19] Ch. 1).
Because developers participate in specification workshops, they take
a much more active part in the whole process. Instead of just passively
reading the documents, they participate in writing the specifications
so they understand them and are much more committed to them.

206

They will understand the specifications correctly

Because acceptance tests provide a focus for development and they

are typically implemented one set at a time, you can be sure that each
and every part of the specification will be thoroughly read and
implemented completely.

They will understand the specifications

correctly
Specification workshops facilitate discussion and allow the developers
to give feedback on the specifications and examples. Realistic examples
leave very little space for misunderstanding. By communicating intent
and discussing goals and anti-goals you make sure that everyone sees
the big picture. Feedback from developers and their participation in
working through examples will give you confidence that they understood you correctly, or reveal areas of conflict that need to be explored
further. No workshop ends until both you and the developers agree
that you understand the same thing. The concrete examples contained
in the set of acceptance tests help make sure that the system will do
exactly what it is supposed to do, significantly reducing opportunities
for ambiguity and misunderstandings.

They will not skip parts of the

specifications
Specifications formulated as executable acceptance tests are impossible
to skip over. If any part of the specification is not implemented, a test
will fail and you'll know it straight away. Even better, developers will
realise this themselves, so you don't need to get involved at all.
Development does not end until all tests are green.
Because tests are based on precise examples, developers have to
implement what is actually specified.

207

Effects on business analysts

You can track development progress

easily
The number of passing acceptance tests is a good realistic measure
of how the project is progressing. It is not a perfect measure since it
does not reflect infrastructure tasks, but as an approximation it is in
practice very reliable. If a phase of development has 70 acceptance
tests and only 20 of them are green, we are still in the first half of the
phase. If you have 65 green tests, you know that you are almost done.
Not all tests are of the same size and complexity, but in general the
differences average out. With an automated test suite, you can easily
run tests at any time and view progress. If a continuous integration
server runs the tests for you overnight, you can get a useful e-mail
report every morning.

You can easily identify conflicts in

business rules
With a readily available set of tests focused on business rules, you can
easily pull out tests related to a particular rule and compare them to
a change request. You do not have to keep all the details of the initial
specification and later changes in your head. Because the tests are
connected directly to the code, if they execute correctly you can be
sure that the specification in the acceptance tests is a reliable and
authoritative statement of what has been implemented. Even if you
do not compare the change request to all the relevant tests, any
conflicting change of the code will break an existing regression test.
Developers will notice a failed test straight away and alert you about
it. This is why you will find about potential conflicts in change requests
and specifications during development, not in production.

208

You'll save time on acceptance and smoke testing

You'll save time on acceptance and

smoke testing
When a reliable set of tests are executed periodically throughout the
project and when each phase of the project is completed with a full,
successful test run, you can be much more confident that the business
rules are implemented in the code. You can easily execute the tests
whenever you want. Because of this, you do not have to verify business
rules manually before a delivery. You can still sanity-test the system
to make sure that some obvious item did not get missed, but you do
not have to spend a lot of time doing it. Knowing that the entire
specification is correctly implemented will give you a lot of confidence
in the code.

Challenges for business analysts

Here are some of the questions and concerns about agile acceptance
testing I hear from business analysts during my talks, workshops and
tutorials.

I don't do tests, that's not my job

Hopefully by now you have realised that acceptance tests are
primarily a live specification of the system. Specifying what the software should do is definitely one of the responsibilities of a business
analyst. If you let developers or testers work out acceptance tests on
their own, you are not ensuring that the effective specification is
correct so you will not be doing your job. If the word test is a stumbling-block to understanding, use one of the alternatives suggested
in the section Better names on page 38.
I find that realistic examples and specification workshops provide a
much better framework for eliciting requirements with customers.
They have to think at a more detailed level to provide concrete answers
to concrete examples and they cannot just brush questions off. This

209

Effects on business analysts

helps significantly to discover the real business rules and to root out
incorrect assumptions, so the customers will be less likely to change
their minds later.
Not only is doing tests part of a business analyst's job, but it makes
the rest of your job much more effective.

Its extra work and I dont have time for

that
Some people tend to keep working the way they feel comfortable with
and pile up more work when they start doing agile acceptance testing.
A business analyst who is comfortable with writing big specifications
documents and lists of requirements might first write them and then
organise a workshop to discuss examples. If such a process helps you
organise your thoughts better or focus on details, I don't see a problem
with it, although you will probably soon find out that the first part
of the process can effectively be skipped. If there is an artefact to be
considered as the physical output of a business analyst's work, then
this should be a set of acceptance tests, not the requirements documents or abstract specifications. Maintaining both acceptance tests
and requirements and synchronising them will take a lot of time and
effort. Out of these, it is really important only to have acceptance tests
up-to-date. So save time and just do acceptance tests.
With the specification workshop as an intensive hands-on laboratory
in which we flush out all the inconsistencies and functionality gaps
and build a shared understanding of the domain, you will actually
spend much less time on transferring knowledge and communication.
If one of your responsibilities is to give the system a quick smoke test
before a release, having a set of automated tests that check business
rules will free up time for this.

210

But they will only look at the tests and not read the requirements...

But they will only look at the tests and

not read the requirements...
Business analysts new to agile acceptance testing often complain about
the fact that when acceptance tests are used as a target for development, developers disregard all other documents and focus only on
the acceptance tests. This complaint comes from a basic misunderstanding about the acceptance tests. Remember that all the business
rules are discussed during the specification workshops, so there should
be at least one acceptance test covering every business rule that needs
to be implemented. As the specification workshops help to build a
shared understanding of the domain, the developers should also have
a lot of tacit knowledge and a much better grasp of the domain without
reading any additional documents.
With agile acceptance testing, the chosen set of acceptance tests
represents the requirements. This is the authoritative source of
information about the domain and the target system. It is kept alive
and relevant throughout the project, so this is what the developers
should be reading. You do not have to write a formal requirements
document unless it helps you focus better, but if you do write one,
keep it to yourself. If you are worried about features not being
implemented because they are not in the acceptance tests, the best
thing to do is to write some new new examples and tests.

What if I leave something out?

A common question I get during introductory workshops from
business analysts is: what happens if something is not specified
precisely in acceptance tests? Realistic and precise examples leave no
space for misinterpretation and they remove ambiguity from the
specifications. The usual reason why features are left out is that the
team is still learning how to implement acceptance testing and can't
get everything done at first, or that there is a lack of information or
insufficient access to domain experts when examples are being written.

211

Effects on business analysts

The fear of these situations is caused by a misunderstanding that there

is a moment when acceptance tests are handed over by the business
analysts to developers and testers, in a similar way to formal
requirements documents being signed off and passed down the
pipeline. There is no such handover in agile acceptance testing. If
anyone is to be the owner of acceptance tests, then this is the business
analyst, and ownership stays with him.
The cost of change with agile development methods is relatively low
compared to more traditional development processes. If your team
is using an agile method, you can always add a new acceptance test
later or change an existing test without causing much trouble. Just
make sure that everyone else knows about the update. If you are
making a large change try to discuss it during a workshop. Changes
are a much bigger problem with a waterfall development process and
this is probably why specification by example and agile acceptance
testing had to wait for agile development methods to mature in order
to gain momentum.
Although they will not openly admit this, some business analysts
intentionally leave ambiguity in formal documents as a sort of a
manoeuvering space for later. This allows them to change their mind
or pass on the blame. Being able to pass on the blame is especially
important in bureaucratic organisations, so ambiguity in requirements
is actually a safety net for these people. I consider that using ambiguity
as a safety net is a very bad and unfair practice that has no place in a
good team. As the cost of later change is not large, there is no need
for the safety net. This approach prepares for the failure of the project,
not for success. If you as a business analyst, with a key role in a project,
are preparing from the start for it to fail, what chance does it have of
success?
In most cases, I find that the fear of blame is unfounded and that we
can be much better off if everyone focuses on doing their job properly.
If you work in an organisation that has a blame culture and you need
to start covering your back even before the development starts, then
you have much bigger organisational problems then those addressed
by agile acceptance testing. If you recognise your own company in

212

I wont understand user stories and tests, Im used to use cases

the previous sentence, instead of slacking and covering your back, it

might be time to polish your resume. At least from what I've seen,
there are plenty of much nicer places to work.
In addition, as agile acceptance testing involves developers and testers
in the process of building the specification, the responsibility to make
it correct and complete is shared. There is less pressure on the business
analysts to get the requirements right on their own.

I wont understand user stories and tests,

Im used to use cases
User stories are not at all difficult to comprehend, and every single
business analyst whom I've worked with and who has made a small
effort to understand the underlying ideas got the picture very quickly.
After all, it is not rocket science, it is just a different approach to the
same problems that people have been dealing with for years. The issue
is often not whether you will understand user stories and story tests,
it is whether you want to understand them. Some people simply resist
any change to the way they work and my only cure for this is to try
and present the benefits as best as I can. This book is my attempt to
do so on a large scale.
There is much excellent literature out there on user stories and I
recommend starting with two books by Mike Cohn, User Stories
Explained[11] and Agile Estimating and Planning[12]. Both are
relatively short so you can read them on the bus or train on your way
to work.
An important thing to remember is that user stories do not have to
be perfect and it's not a problem if you don't get them right straight
away. They are there primarily as a collaborative planning tool and
to stimulate discussion. When you are using them, other people will
be involved as well, helping you understand how to use them and to
find better ways to exploit user stories. The situation with story tests
is very similar. They still contain triggers, steps, exceptions and all
the usual utilities in a business analyst's toolbox they are just written

213

Effects on business analysts

down differently. Because tests are produced during the specification

workshops, the other people will help you produce more complete
tests and specifications. Developers will be there to help you write
them down in the particular way required by your team's automation
tool of choice.

Its impossible to describe non-functional

requirements as tests
User interfaces, performance, usability and security requirements are
very hard to describe effectively as automated acceptance tests. This
is very often used as an argument against moving towards agile
acceptance testing from a more traditional requirements and
specification process. My counter-argument for this is that the biggest
benefit of agile acceptance testing is in the conversation and building
a shared understanding based around realistic examples. The fact
that the user interface cannot be described as a FIT table or a BDD
scenario does not prevent you from discussing the interface and
providing examples during the specification workshop. Not everything
can be automated, so what? You can find some ideas on how to deal
with this issue in the section What do we do with things that cannot
be automated? on page 90.
There is nothing preventing you from using a different tool or
processes in addition to agile acceptance testing. The practice is not
a magic wand that will make all your problems go away, but it will
help you handle functional requirements much more effectively than
with traditional processes. You can organise workshops with security
experts and usability designers to help developers produce software
that meets security and UI requirements. You can also schedule
security probing, performance testing and usability reviews periodically every few weeks or months in addition to the functional test
automation.

214

There is no big picture

There is no big picture

Breaking business rules down into lots of realistic examples to illustrate all the variations in possibilities might make it hard to see the
big picture. Business rules may become too scattered throughout the
acceptance test sets and if you do not take care, they will become hard
to understand after a while. However, most acceptance test tools allow
you to group tests into hierarchies and add arbitrary descriptions to
tests or groups of tests. Always group related examples together and
provide descriptions of the relevant business rules to explain examples
better. I've outlined some ideas on how to organise tests efficiently
in the section Organise tests so that they are easy to find on page 134.

There is no traceability
When a single person acts as a gateway for all the requirements and
everything is put into a single document, it is relatively easy to keep
track of who ordered what and when, especially if the document is
signed off before development starts. Because of the collaborative
model of agile acceptance testing and because a single phase of
development can be broken down into hundreds of small examples
and tests, some people are concerned about the traceability of
requirements and specifications. After all, implementing a requirement costs money and someone ultimately has to approve and pay
for it.
I don't like sign-offs and prefer to work closely with the customers
on building the system that they need. However, if you need to get
official approval, treat the formal set of acceptance tests as the
specification and get a sign-off on it. Most test automation and
management tools allow you to put arbitrary comments on tests, so
you can include the name of the person who approved a particular
test and the date when it was approved. If you implement strict
ownership, which anyway I think is a bad idea, you can list the name
of the person who owns a particular test as well.

215

Effects on business analysts

When user stories are used for project planning, acceptance tests are
derived from stories and you can implement traceability by linking
individual tests to a story, then linking stories to customer benefits
and project goals. You can track who requested a story and when on
individual story cards, or you can put this information into an
acceptance test hierarchy.
If the tests are stored in a version control system, as I recommend in
the section Keep tests in a version control system on page 135, then it
is also very easy to go back and inspect who changed a test, when and
why. Most source version control systems give you this functionality
out of the box.

Stuff to remember
The role of the business analyst is changing from that
of a conveyor of information to that of a facilitator who
enables people to share knowledge directly.
The business analyst is a great choice for the facilitator
of the specification workshop.
Working with acceptance tests makes the rest of the
business analyst's work easier and more effective.
You can still keep your traditional documents, but use
acceptance tests as the authoritative specification.
You can modify the specification (acceptance tests) when
the development starts without causing great problems
for anyone.
There should be no handover of tests from business
analysts to developers or testers.
Tests don't have to be perfect from the start.
Getting it right is not your sole responsibility. Developers
and testers share responsibility for the specifications.

216

Chapter 13

Effects on testers
In 2005 I worked with a development organisation that had a remote
testing department. The testing team was considered completely
ineffective since many problems passed undetected through them,
and developers generally considered talking to the testing department
as a waste of time. To be fair to the testers, they were expected to grasp
instantly whatever the developers produced and test it without
knowing the first thing about it. Nobody wanted to spend any time
explaining anything to them and there was simply no feedback in the
whole process. I think that having a detached testing team simply
does not work, nor does getting testers involved at the end. In the
words of Philip Crosby, Quality has to be caused, not controlled.
This does not mean that there is no place for testers in agile processes
far from it but that the role of the tester has to change. Agile
acceptance testing is a great manifestation of this.

The new role of the tester

Mary Poppendieck talked about her experiences of moving to a lean
production process at Google Tech Talks on the 15th of December
2006, in a session titled Competing on the Basis of Speed.1 Although
her experiences relate to manufacturing, I think that they perfectly
describe what the role of the tester in software development should
be and why this is important. She said:
When we started up in our plant, we had people in QA who
used to try to find defects in our products, and we moved them
all out on to the production line to figure out how to make stuff
without defects in the first place...You will be amazed at how
much faster you go when you make stuff and defects are caught
when they occur instead of being found later.
1

http://video.google.com/videoplay?docid=-5105910452864283694

217

Effects on testers

Instead of acting as controllers at the end of the project, testers should

help developers and business analysts avoid problems. Agile acceptance testing allows testers to suggest how a piece of software should
be verified, rather than manually test the software themselves. Testers
need to communicate concerns and possible pitfalls from the testing
perspective during specification workshops and warn business people
and developers about important edge cases that they have missed.
By taking an active participation in the specification workshops,
testers learn about the domain and the expected system behaviour
first hand. This solves the problem of testers not really knowing
anything about the system when the time comes to approve a release.
The knowledge gained during the workshop will be instrumental in
writing, expanding and refining tests.

Benefits for testers

As a quick recap of Part II, let's go back to the benefits promised in
the introduction and see how the practices of agile acceptance testing
deliver them.

You can stop developers from making the

same mistakes over and over
If you've had some experience working with the development team,
you surely know what kind of mistakes they are likely to make. Instead
of checking manually for the same problems over and over again and
returning the release back to the developers because of these issues,
you can prevent the problems from happening in the first place. The
specification workshops allow you to suggest these typical problems
as examples that should be discussed and put into the acceptance test
suite. Discussing the examples with developers helps them learn and
helps prevent similar mistakes in the future. Including tests for typical
problems in the acceptance test suite ensures that developers sort out
these issues before submitting the system for testing in the first place.

218

You will have a much better understanding of the domain

As you can add tests for any bugs found in production to the acceptance test suite, you can be sure that the bugs will not resurface, or at
least that the developers will find out about resurrected bugs and fix
them before you need to get involved.

You will have a much better

understanding of the domain
If you participate in the specification workshops and learn about the
goals of iterations you will understand the system at a much deeper
level than if you had simply read specifications after development.
Even better, the process of building a shared understanding during
the workshop ensures that your ideas about the system are consistent
with what the customers want and what other members of the
implementation team understand.

You'll delegate a lot of dull work to

developers
Agile acceptance testing relies heavily on the automation of tests and
you will benefit from this. There will be a solid foundation for the
automation of most of the QA process, so that you do not have to
perform tests manually. Even better, instead of you automating the
tests on your own, this responsibility is either given completely to
developers or shared with them. Your job is to participate in
proposing the tests that should be automated. So you save a lot of
time because you won't have to run tests manually or spend time
automating them. You will have much more time for verifying things
that cannot be automated easily and other important tasks.

You can build in quality from the start

Instead of having a reactive role, trying to catch problems in a
developed system, agile acceptance testing puts you in the driving
seat and allows you to avoid problems. You can build quality into the

219

Effects on testers

software from the start by suggesting examples that would break or

cheat the system during the specification workshops. The ensuing
discussion will help you to teach developers how to think more like
you and what they need to do to make the product better.

You'll be able to verify business rules at

the touch of a button
Agile acceptance testing incrementally produces a set of regression
tests as the software is being implemented. This set of tests is automated and you can execute it at the touch of a button, allowing you
to verify and reverify the business rules implemented in the system.
Because developers automate tests to connect directly to business
code, not through the user interface, tests run much more quickly
than if you automated them using one of the classic UI test robots. I
don't mean to say that having automated UI tests using robots is bad,
by all means use them when it makes sense. But you don't have to
verify business rules with such tests automated acceptance tests
provide a much more efficient way to do this.

You will have a lot more time for

experimenting
Because automated acceptance tests already give you a level of
confidence in business rules, you will have a lot more time for experimenting with the system. Because you are involved in writing
examples, building the specification and creating acceptance tests,
you will have a much better understanding of which areas have not
been considered or covered in enough detail, so that you can focus
your exploratory testing effort more accurately.
This is one of the key contributions that testers should make to the
project. Developers and business analysts think about the system from
the perspective of making it work and they usually don't try to use
the system in unexpected ways or go down unpredictable paths.

220

Better relationships in the team

Testers, on the other hand, are specialised in this. By giving you more
time to do this, agile acceptance testing empowers you to perform
better exploratory testing and provide much better feedback from
your experiments.
Having more time for experimenting also allows you to refine and
expand the suite of regression tests based on what you learn during
exploratory testing and to suggest more possible problems in workshops for iterations to come. So acceptance testing, test automation
and exploratory testing are linked together in a positive reinforcing
loop.

Better relationships in the team

Agile acceptance testing also helps to build better teams and bring
people together. When the primary role of testers is to find defects,
they are often seen as adversaries of developers. Developers and testers
can consider each other as rivals. With agile acceptance testing,
nobody has to be a bad guy. The specification workshops require
people of different profiles to work together and helps testers build
in quality from the start without getting into conflict with developers.
From my experience, I can state emphatically that bringing people
together to discuss the examples during the workshops also helps to
build trust, mutual respect and personal understanding among the
participants, transforming disparate groups of individuals into teams.
By working closely together with developers, you will learn more
about the system architecture and implementation and you will be
able to provide better feedback. Developers will start to respect your
input much more than if you were working separately. As testers are
required to participate during the specification workshops, your
opinions will matter and business people will notice this and listen
to you.

221

Effects on testers

Challenges for testers

Here are some of the questions and concerns about agile acceptance
testing I hear from testers during my talks, workshops and tutorials.

I don't want to lose control over testing

Testing managers, especially those in big corporate organisations,
often look at agile acceptance testing as another attempt by developers
to take over part of their territory and resist this change with all the
power they have. In my view, this is plain silly and always has to do
much more with corporate politics than reality. If anything, agile
acceptance testing gives testers more responsibility and control. It
doesn't stop you doing any kind of verification that you were planning
to do, but it does transfer much of the more boring manual testing
onto the shoulders of developers, who then automate most or all of
it. So you are not giving away control over testing, you are obtaining
tools to do it more efficiently and involving more people in testing.
Agile acceptance testing gives testers much more influence over the
project the testing department can finally have a say in what gets
produced, not just what gets fixed. It also allows you to be more
productive in testing, because you will run into fewer blocking issues
as key concerns will have been communicated before development
starts.

It's cheating
A very strange argument I have heard on a few occasions against agile
acceptance testing is that it is effectively cheating. The argument goes
that with agile acceptance testing, testers are required to tell the
developers exactly what they are going to check, so the developers can
implement the software to pass the tests upfront.
I honestly struggle to understand why this is a problem. Software
development, at least in my eyes, is a cooperative activity and not a

222

Will I lose my job because of this automation?

competition. If the developers don't put the problem into the system
in the first place, it's more efficient than you catching the problem
later and returning it back to them for fixing. It does not matter who
put the bug in, who caught it and who fixed it. Agile acceptance testing
asks us to not create the bug in the first place.
Unfortunately, agile acceptance testing has no chance of success in
companies where people earn money on metrics such as bugs found
and bugs resolved. If your salary or bonus depend on such statistics,
you are not going to like agile acceptance testing one bit. On the other
hand, if you consider improving the quality of the software as your
job, not catching bugs, then this practice can help you greatly to do
it more efficiently.
The idea of giving developers tests upfront is similar to the idea of
publicising where speed cameras are on the motorway. The goal of
introducing speed cameras is to reduce accidents, not to charge people
for speeding. If you tell everyone where the cameras are and if there
are enough cameras on the most accident-prone parts of the road
system, people almost always stop speeding there. Testers on an agile
team should ensure that there are enough speed cameras and that
they are placed on all parts of the road that are likely to cause accidents.

Will I lose my job because of this

automation?
Some testers whom I've spoken to about this practice fear that they
will be made redundant by test automation. Test automation, for
things that can be automated, is coming regardless of agile acceptance
testing. It is going to happen whether testers are on board or not. I
really think that the days of manual testing scripts written down on
paper and carefully executed by twenty testers over two weeks are
numbered. This kind of testing will never be completely rooted out,
just as there are still jobs for COBOL programmers, but most testing
activity in the future will not revolve around manual execution of test
scripts.

223

Effects on testers

A very important thing to realise is that test automation does not

replace human testers, it helps them do their work better and leaves
more time for things that require human intelligence and touch. This
includes manual exploratory testing, looking for inconsistencies,
testing with more extreme conditions, performance, security and
usability testing. Instead of swimming against the current, use the
change to your advantage. Partial automation even helps testers to
do exploratory testing better, by by speeding up the boring parts.
I would argue that, rather than endangering testers' jobs, agile
acceptance testing increases their importance to the company and
make their jobs more secure. It offers several options for career
advancement, including becoming an automation specialist or a
domain specialist.
A tester with domain knowledge is absolutely irreplaceable, even with
a totally agile development process and aggressive test automation.
I learned this the hard way, while working on a new trading application for a big online betting system. We did everything by the book,
had good customer involvement, business analysts who participated
in the project and a development process that could be considered
agile by most standards. Because testing was outsourced and inefficient, we decided to ignore the testing department and rely on automated tests that we did ourselves, although we still had to hand over
the system for formal testing eventually. The outsourced testing
department never found anything really serious. I later realised that
this was because they had no idea what they were testing. No-one was
interested in sharing domain knowledge. Before the first release, the
company hired a senior tester from one of their clients. This woman
had worked on similar systems for a long time and knew the business
domain inside and out. Coming from a QA perspective, she had
excellent ideas about how to break the system and knew exactly where
it might be vulnerable. We agreed to let her have a go at our system
just for fun, confident that automated unit and acceptance tests
protected us from problems. It was a very painful experience to watch
her tear the application apart. She did not find any serious problems
in business rules, but she found loads of small embarrassing issues
in the first ten minutes of smoke testing. I learned to respect the skills

224

Will I lose my job because of this automation?

and experience of testers as very important additions to the skills and

insights of developers and business people.
I believe that testers have a completely different view of software
systems than developers or business analysts. They think about
breaking and cheating the system, in addition to verifying that it
works. Although developers are typically good at identifying edge
cases in business rules, they rarely check the user interface properly
and they don't approach the system with the intention of breaking it.
This is exactly why testers need to participate in the specification
workshops and point out these cases. Developers and business analysts
are often simply blind when it comes to this.
If you are worried about your job, learn all you can about the business
domain. With a testing mind-set and business knowledge, you can
help developers and business people see their blind spots. This will
make your contribution much more visible and your position much
more important. Specification workshops help you do this because
you will be right there with the experts, getting the chance to learn
about the domain and discuss the software before it gets developed.
Mike Scott wrote2 that he had noticed the blending of analyst and
tester roles: when we start recognising the tests as executable
specifications you rapidly realise that the role is practically the same.
He gives the example of a client organisation that makes no distinction
between testers and business analysts, and just has a single analyst
role. According to Scott the only problems with this are personal, as
business analysts tend to think of a test role as a demotion. I noticed
the same thing in an organisation that I provided consultancy to in
early 2008, where one of the business analysts was really enthusiastic
about acceptance testing but her superior rejected the idea. The analyst
was given permission to participate in acceptance testing provided
her role was renamed to test analyst. Her boss insisted that she was
no longer a business analyst if she was involved in anything related
to testing.

in a private e-mail

225

Effects on testers

Antony Marcano suggested a career path for testers who have an

inclination for programming and developers who demonstrate an
aptitude for testing, during his talk on Understanding QA/Testing on
Agile Projects3 in September 2008 in London. By growing their skills
in these areas, they may become what Marcano calls tester-developers
or developer-testers. Microsoft is also promoting this idea,4 calling
this role software design engineer in test (SDET). An SDET has the
following responsibilities: testing software components and interfaces
in more technical depth, writing test programs to assure quality and
developing test tools to increase effectiveness. Marcano thinks5 that
this will soon become the norm for testers and developers alike and
draws a parallel between developer-testers and the blending of analyst
and programmer roles in earlier times. The role of analyst-programmer was introduced because developers were increasingly expected
to perform systems analysis. Analysis was then taken as one of the
responsibilities of programmers and the analyst-programmer title
was later simplified to just developer.
Another very interesting career option is to become an automation
specialist and learn how to bridge the gap between examples coming
out of the workshop and domain code. In general, automation
specialists are paid more than click-testers and there seems to be a
healthy demand for people that know their way around Fitnesse,
Selenium and similar tools at the moment. This career path might be
especially interesting for technical people who know a bit about
programming and possibly came to QA from a programming background. Automating tests requires some basic programming skills,
but it does not demand that you know all about the latest frameworks
and technologies.
In any case, agile acceptance testing practices can help you on this
path as well, as it encourages you to work closely with developers.
Pair up with a developer and learn how they write fixtures and automate tests, then start taking over these parts of the role.

http://skillsmatter.com/podcast/home/understanding-qatesting-on-agile-projects
http://members.microsoft.com/careers/careerpath/technical/softwaretesting.mspx
5
http://www.testingreflections.com/node/view/7280
4

226

Acceptance tests will not catch all problems

Acceptance tests will not catch all

problems
Agile acceptance testing is not a silver bullet and it would be silly to
expect that it will magically catch all problems. Again, the focus of
the practice is much more on building a shared understanding of the
domain and guiding the development then on quality control.
Some things cannot be easily automated and every tool that you decide
to use will have its good and bad sides. But applying agile acceptance
testing does not prevent you from running any other tests or using
different tools. It even gives you more time to focus on things that
cannot be caught by acceptance tests. Instead of verifying every single
business rule, you can be confident that they are correct and spend
more time sanity testing or running exploratory tests.
User interface validations are not likely to be fully automated in the
immediate future. This is where testers will still need to utilise their
skills. But instead of verifying business rules, which should be covered
by acceptance tests, you can just focus on checking that the actual
user interface is solid and that it works as expected.
Also, while a team is learning how to use agile acceptance testing and
getting buy-in from their business users or customers, it is likely that
some things will still fall though the communication gap. Instead of
using this to point out flaws in agile acceptance testing, I urge you to
work out why things were missed in the first place and help improve
the specification workshop for the next iteration. As a tester, your
job is to come up with bug prevention strategies, not to find bugs at
the end.

Stuff to remember
The new primary role of testers is to help people avoid
problems, not to discover them.

227

Effects on testers

The QA process does not end with agile acceptance

testing. You can still use any other tool or technique in
addition to acceptance tests.
The automation of acceptance tests gives you more time
to test things that cannot be easily automated, where
human intelligence and touch are important.
Testers have more clout in projects that use agile
acceptance testing.
Develop your career by learning all you can about the
domain during specification workshops, or become an
automation specialist.

228

Chapter 14

Effects on developers
Although the role of developers changes with the introduction of agile
acceptance testing, I find that they are generally enthusiastic about
the idea and do not put up any resistance. From a technical
perspective, most developers see it just as an extension of test-driven
development principles, and TDD is now becoming well established
among software developers. However developers are sometimes overly
enthusiastic and forget about the requirement to collaborate. They
take over more responsibility than they should, losing the benefits of
shared understanding and communication.

New responsibilities for developers

Agile acceptance testing requires developers to take a more active
part in the process of eliciting requirements and building a specification of what is to be developed. Many senior developers already do
this, especially on agile teams, so this is just a reinforcement of the
requirement. If your team leaves the creation of specifications entirely
to business analysts and customers, then it is time to get more
involved.
A major change introduced by agile acceptance testing is the
requirement for developers to participate in domain discussions,
evolving their own understanding of the domain and providing much
needed technical expertise during the specification workshops.
Although this duty has nothing to do with writing code, you should
take it very seriously as it will help you write the code much more
effectively later. While the examples are being worked out, it is the
role of developers to think ahead about implementation challenges,
identifying functional gaps and technical edge cases, suggesting them
for discussion and explaining to business people why these cases
matter. Business people often have a very focused perspective of the
system, concentrating on how it should behave when it works correctly
229

Effects on developers

and ignoring things they consider minor problems. These same

problems could be really significant for developers, because it is their
necks that are on the line when something does not work. A former
colleague of mine whined about the fact that the business people he
worked for refused to discuss a potential problem because the chances
of it happening were only one in a million. From the perspective of
the business people, the problem was irrelevant. On the other hand,
the system they were supporting processed about 200,000 business
transactions a day, so for my colleague one in a million meant that
he would be called in at least once a week to troubleshoot serious data
consistency problems. Developers need to raise awareness about such
concerns and get them properly discussed and resolved during
specification workshops.

Benefits for developers

As a quick recap of Part II, let's go back to the promises made in the
introduction and see how agile acceptance testing makes them come
true.

Functional gaps will be flushed out before

development
With the focus on realistic examples and writing them down as formal
acceptance tests, it is easy to spot inconsistencies and gaps in the
requirements during specification workshops. Writing down examples
has the same effect as writing code, making it obvious when a particular case is not covered. While the examples are being discussed and
written down, developers should try to think about implementation
challenges and possible edge cases that also need to be discussed. Once
the gaps and edge cases are identified, they should be discussed and
resolved while the business people are in the room and focused on
the problem. Using this process, we collaboratively produce a development specification that is much more detailed, complete and
consistent than traditional requirements specifications. Gaps should

230

Business analysts will really understand special cases

no longer be identified only once we start developing code, when the

business people might no longer be readily available to discuss them.

Business analysts will really understand

special cases
If you are unclear about a particular example or if you suspect that
there is an important edge case that needs to be covered, you can offer
it for discussion during a specification workshop. If business people
seem unconcerned with the example, explain first why it is important
and what difference it makes. Examples based on precise realistic
information require people to think harder and offer precise results,
so they will not be able to just brush the question off. Go through the
example and work out together whether it is important enough to be
added to the specification. The primary goal of the specification
workshop is to build a shared understanding of the domain, so a
discussion on an edge case should end either by everyone agreeing
that the example is not important and knowing why it is not important, or everyone agreeing that it is important and working out how
the system should behave in this case.

Automated tests will be your targets for

development
Acceptance tests provide us with a universally agreed target for
development. They should cover all important cases (or at least most
of them) so that we can focus on implementing only what is specified
in these cases. Examples that were not selected for the set of acceptance
tests are not relevant for the project or not important for the current
phase of development, so we do not have to think about implementing
code just in case.
Because tests are automated, it is easy to check whether the code
implements the specification correctly and completely. A failing test
suggests that we need to implement additional functionality. When

231

Effects on developers

all the tests pass, we know that the work is done. If a previously passing
test fails, we have broken something.

Code will be easier to share, hand over

and take over
An acceptance test serves not only as a specification of what the code
should do before it is developed, but also as a description of what the
code does after implementation. Because tests are executable and kept
in synchronisation with the code, this description is accurate and
authoritative. Acceptance tests are much more concise than code, so
they are a lot easier to read and understand.
Having a set of acceptance tests to describe what a code module does
makes it much easier to share it with someone else. They will not have
to read through the code or depend on you to describe it. The first
thing I do typically when new members join my team is to give them
a brief introduction to the project and then let them read existing
acceptance tests to get to know the system better.
Having this kind of live documentation is also very useful when you
need to hand over the code to another team or organisation.

Challenges for developers

Here are some of the questions and concerns about agile acceptance
testing I hear from developers during my talks, workshops and
tutorials.

I don't have time to write fixtures or

maintain tests
Technically, agile acceptance testing introduces new work for
developers, but not necessarily more work. By new work I mean that
developers have to participate in writing tests, write the automation

232

Acceptance tests turn out to be incomplete

code and maintain it. On the other hand, this should produce a much
more accurate and clear target for development, reducing later rework
caused by misunderstandings or incomplete specifications. From my
experience, the overall result is much less work for developers rather
then more work, although in the beginning this new work does seem
to be additional.
One potential issue with this is that the benefits of agile acceptance
testing are not that obvious or instant. You won't necessarily notice
straightaway that you have less rework to do, since there is usually a
time lag between delivery of code and requests for rework arriving.
The reductions in bug numbers and change requests are also hard to
spot when you are looking only at the current iteration. To truly
evaluate the benefits of agile acceptance testing, you have to look at
the way you work from a wider perspective and compare the whole
process after the introduction of acceptance testing with the way it
was before.

Acceptance tests turn out to be

incomplete
Ideally, we want to flush out all edge cases and discover all functional
gaps in the specification workshops, and agile acceptance testing
normally helps greatly in this direction. However, sometimes things
slip through the net, especially when you are just starting to use agile
acceptance testing. If you start developing code and run into an
unidentified case, grab a business analyst or customer representative
and discuss it. Don't just go ahead and assume that you know how
the system should work. You should also discuss why the case was
missed during the workshop and whether any other rules are affected
by it. When you discuss the case, take care not to over-generalise and
go beyond the scope of the current iteration. Things that were intentionally left out of scope should not creep back into the system.
The discovery of many new edge cases in development is typically a
signal that the specification workshop didn't function properly or
didn't take place at all. The workshop should not be a one-way event

233

Effects on developers

where the customer or a business analyst presents his ideas and

developers who then go away and implement them. A workshop like
this will have the same consequences as the orders of battalion
commanders as described in the section Imperative requirements are
very easy to misunderstand on page 8. An effective workshop requires
the active participation of all participants. Only then can we build a
shared understanding of the domain.
Participating in the specification workshops is not a waste of time, it
is the best chance to learn about the domain and the right time to
identify and discuss functional gaps. It should be considered as one
of the key responsibilities of programmers, not as something unimportant that can be done on a voluntary basis.

Can't we just write the tests ourselves?

Developers often see acceptance tests more as larger unit tests then
as a communication technique, and therein lies one of the biggest
pitfalls of agile acceptance testing.
While writing unit tests before code is recommended, sometimes
writing code and unit tests in parallel works fine as well. If I am unsure
about a particular case after I have written the code, I can quickly
write a unit test to verify it. This does not work for acceptance tests,
or at least not in the same way. With unit tests, developers own the
code and they should perfectly understand how it is intended to work.
This is why writing a unit test in parallel with the code is not a problem. However, acceptance tests must always be written before the
code, because they specify what the system should do. By adding an
acceptance test, you are extending the specification of the system and
making a business decision about the domain. You are then effectively
taking on the role of the customer.
Any additional tests or changes to existing tests must be discussed
with business people, just in case there is a mismatch in what you
think it should do and what they think it should do. As you gain more
knowledge about the domain, you will make better and better guesses,
but no guess should go unverified.

234

There is too much duplication

The worst thing developers can do is to write acceptance tests on their

own just to have good code coverage or to tick a box that the project
has acceptance tests. Writing acceptance tests after the code just to
pass a test-coverage criterion is pointless, because you are writing a
specification based on what was already developed. The only exception
to this that I can think of is where you build a set of tests to document
what a legacy system does before it is improved or modified.
Rather than waiting for a phase of the project to end and then
discovering parts of the code that are not covered by tests, chase your
business people to help you with defining how the system should
behave before developing these parts. Once you have some experience,
you will identify most of these cases in the specification workshop
anyway, so you should not have such a big problem with code test
coverage.
If there is no buy-in from the business side and they do not want to
discuss examples, I would suggest that you are better off without the
acceptance tests than with acceptance tests that you make up
yourselves. At least you will not have a misleading specification.
This does not mean that developers cannot write functional tests at
all. Unit tests, component tests and integration tests can and should
all be written to reflect the developer's view of the system and concerns
with a particular piece of code or way of working. They might even
be automated using the same tools as acceptance tests, but they are
not a replacement for them, and neither do they in any way make up
for the communication benefits obtained through specification
workshops.

There is too much duplication

With acceptance tests and unit tests in place, developers often notice
duplication, which they typically don't like. Duplication in code is
one thing, and we are right to fight against it relentlessly, but duplication in tests is something else and is not such a bad thing. Unit tests
and acceptance tests serve completely different purposes. Unit tests
verify technical correctness from the developer's perspective.

235

Effects on developers

Acceptance tests verify functional correctness from the customer's

perspective. The same piece of code can be subjected to both verifications separately. Having said this, once acceptance tests come into
play, you can really focus unit tests only on technical correctness and
not write any business rule unit tests. This takes some time to get used
to, but it is a good way to remove some of the duplication. See the
section Unit tests vs acceptance tests on page 113 for an example.
Acceptance tests are written based on examples developed during the
specification workshops, so there will often be some duplication even
in the examples or tests themselves. Again, this is not necessarily a
bad thing. If the system is easier to comprehend and explain with a
bit of duplication, leave it in. On the other hand, duplication in
specifications and acceptance tests might also suggest that we are
missing an important abstraction that should be made explicit to
make the system easier to understand.
Some acceptance test automation tools allow you to extract common
parts of tests into a shared set-up or separate components that can
be included in other tests. As a developer, it is your job to suggest
such changes so that tests become easier to maintain.
Resist the urge to restructure and reorganise tests so that they are
easy to maintain but no longer easy to understand. Acceptance tests
are not code. They need to be used for communication with business
people over the life of the project, so they have to be able to read and
understand them. If you reformat a test so that only a developer can
understand it, you might as well throw it away.

I cannot run acceptance tests from my

IDE
Agile acceptance testing tools are designed for collaboration rather
than for efficient development, for business users rather than for
developers. Comparing acceptance tests to unit tests again, this situation makes life more difficult for developers. Most integrated
development environments now allow you to run unit tests directly

236

Unit tests give us better feedback

from the code-editing window, shortening the loop between developing and verifying a piece of code. Acceptance tests cover much larger
parts of the code, so it is not so easy to work out exactly what they
relate to, and current tools often lack good IDE integration. One of
the biggest complaints is that it is often much harder to debug code
under acceptance tests than with a unit test. The biggest benefit of
agile acceptance testing is improved communication and collaboration. This easily outweighs the importance of close integration and
easy debugging during test runs.
The only thing I can offer as a consolation is that tools are evolving.
One of the most important areas where tools need to improve is to
provide multiple views for different people, so that developers can
keep working from an IDE and business people can use something
more convenient for them. For the time being, tools are a big problem
for everyone, but the point of agile acceptance testing is not in tools
but in the conversation.

Unit tests give us better feedback

Acceptance tests are often compared to unit tests by developers,
although they do not solve the same problems. Acceptance tests deal
with larger pieces of code and longer development efforts, so you
don't get the satisfaction of instant feedback provided by unit tests.
Slower feedback and the pain of maintaining the glue between
acceptance test descriptions and domain code can cause disappointment among developers with acceptance testing, because they do not
see any quick benefits. A common complaint from developers is that
unit tests give them much better and faster feedback on the code.
You need to understand that acceptance tests are not really there to
provide feedback on the code, so expecting them to do so is a mistake.
Acceptance tests are primarily a communication tool for people.
Comparing them to unit tests, even from a developer perspective,
does not make any sense. Use normal unit tests to drive the development of small pieces of code, use acceptance tests to guide the overall
effort.

237

Effects on developers

Stuff to remember
You cannot write acceptance tests yourself.
You need to participate in domain discussions.
Specification workshops give you a good chance to
discuss edge cases and inconsistencies with domain
experts before development.
Acceptance tests do not provide instant feedback as unit
tests do. This is because they deal with the overall picture
and business rules, not code units.
You need both unit tests and acceptance tests.
Acceptance tests exist to facilitate communication
between people.

238

Appendix A.

Resources
Books and articles
[1] Tracy Reppert. Copyright 2004. Software Quality Engineering.
Better Software Magazine: Don't just break software, make
software. July/August 2004.
[2] Frederick P. Brooks. Copyright 1995. Addison-Wesley Professional. The Mythical Man-Month: Essays on Software Engineering, Anniversary Edition (2nd Edition). 0201835959.
[3] Gerald M. Weinberg and Donald C. Gause. Copyright 1989.
Dorset House Publishing Company. Exploring Requirements:
Quality Before Design. 0932633137.
[4] Gary Klein. Copyright 1999. The MIT Press. Sources of Power:
How People Make Decisions. 0262611465.
[5] Bertrand Russell. Copyright 1985. Open Court Publishing
Company. The Philosophy of Logical Atomism. 0875484433.
[6] James Surowiecki. Copyright 2005. Addison-Wesley Professional. The Wisdom of Crowds. 0385721706.
[7] Tom Davenport. Copyright 2003. CXO Media Inc. CIO
Magazine: A Measurable Proposal. June 2003.
[8] Robert C. Martin and Grigori Melnik. Copyright 2008. IEEE
Software: Tests and Requirements, Requirements and Tests: a
Mobius Strip. January/February 2008. page 54.
[9] Rick Mugridge and Ward Cunningham. Copyright 2005.
Prentice Hall PTR. Fit for Developing Software: Framework
for Integrated Tests. 978-0321269348.

239

Resources

[10] David Lorge Parnas, Ryszard Janicki, and Jeffery Zucker. Copyright 1996. Springer Verlag. Relational Methods in
Computer Science:Tabular representations in relational
documents . pages 184-196. 978-3-211-82971-4.
[11] Mike Cohn. Copyright 2004. Addison-Wesley Professional.
User Stories Applied: For Agile Software Development. 9780321205681.
[12] Mike Cohn. Copyright 2005. Prentice Hall PTR. Agile Estimating and Planning (Robert C. Martin Series). 0131479415.
[13] Eric Evans. Copyright 2003. Addison-Wesley Professional.
Domain-Driven Design: Tackling Complexity in the Heart of
Software. 0321125215.
[14] Karl Weick. Copyright 1984. Jossey-Bass. The Executive Mind:
Managerial thought in the context of action. 0875895840.
pages 221-242.
[15] Shigeo Shingo. Copyright 1986. Productivity Press. Zero
Quality Control: Source Inspection and the Poka-Yoke System.
0915299070.
[16] Mary Poppendieck and Tom Poppendieck. Copyright 2003.
Addison-Wesley Professional. Lean Software Development:
An Agile Toolkit. 0321150783.
[17] George L. Kelling and James Q. Wilson. Copyright 1982. The
Atlantic Monthly Group. Atlantic Monthly: Broken Windows.
March 1982.
[18] Andrew Hunt. David Thomas. Copyright 1999. AddisonWesley Professional. Pragmatic Programmer: From Journeyman to Master. 978-0201616224.
[19] Ellen Gottesdiener. Copyright 2002. Addison-Wesley Professional. Requirements by Collaboration: Workshops for Defining
Needs. 0201786060.

240

Online resources

[20] Steve McConnell. Copyright 2000. IEEE Software: Cargo Cult

Software Engineering. March/April 2000.
[21] Mary Poppendieck and Tom Poppendieck. Copyright 2006.
Addison-Wesley Publishing Company. Implementing Lean
Software Development: From Concept to Cash. 0321437381.
[22] Gojko Adzic. Copyright 2008. Neuri Limited. Test Driven .NET
Development With FitNesse. 0955683602.
[23] Jennitta Andrea. Copyright 2007. IEEE Software: Envisioning
the Next Generation of Functional Testing Tools. May/June
2007.

Online resources
Here are the links to all the online resources mentioned in the book.
You can find all these links and more on the accompanying web site
http://www.acceptancetesting.info.

Talks and videos

Adam Geras: Virtual Machines to support testing from Agile 2008.
http://video.google.com/videoplay?docid=-6298610650460170080
Antony Marcano: Understanding QA/Testing on Agile Projects from
Skills Matter Agile Month in 2008. http://skillsmatter.com/podcast/home/understanding-qatesting-on-agile-projects
Elisabeth Hendrickson: Things that change from Agile 2008.
http://video.google.com/videoplay?docid=8565239121902737883
Gojko Adzic: Introduction to Agile Acceptance Testing from Skills
Matter Agile Month in 2008. http://gojko.net/2008/09/19/agileacceptance-testing-video/
Martin Fowler and Dan North The Yawning Crevasse of Doom from
QCon London 2007. http://www.infoq.com/presentations/FowlerNorth-Crevasse-of-Doom

241

Resources

Mary Poppendieck: Competing on the basis of speed from Google

Tech Talks in 2006. http://video.google.com/videoplay?docid=5105910452864283694
Rick Mugridge: Doubling the value of automated tests from Google
Tech Talks in London in 2006. http://video.google.co.uk/videoplay?docid=-7227306990557696708

Presentations
Gilles Mantel: Test Driven Requirements Workshop from Agile
2008. http://testdriveninformation.blogspot.com/2008/08/materialof-tdr-workshop-at-agile-2008.html
Michael Phoenix: Domain Specific Testing Languages from Agile
2008.
http://www.solutionsiq.com/agile2008/agile-2008domain.php
Naresh Jain: Acceptance Test Driven Development.
http://www.slideshare.net/nashjain/acceptance-test-driven-development-350264

Articles
Brian Marick: An Alternative to Business-Facing TDD.
http://www.exampler.com/blog/2008/03/23/an-alternative-tobusiness-facing-tdd/
Brian
Marick:
My
Agile
testing
project.
http://www.exampler.com/old-blog/2003/08/21/
Dan North: Introducing BDD. http://dannorth.net/introducingbdd
Gojko Adzic: Fitting Agile Acceptance Testing into the development
process. http://gojko.net/2008/09/17/fitting-agile-acceptance-testing-into-the-development-process/
Jennitta Andrea: Envisioning the next generation of functional
testing
tools.
http://www.jennittaandrea.com/wpcontent/uploads/2007/04/envisioningthenextgenerationoffunctionaltestingtools_ieeesw_may2007.pdf

242

Tools

Jim Shore: How I Use Fit. http://jamesshore.com/Blog/How-I-UseFit.html

Robert C. Martin and Grigori Melnik: Tests and Requirements,
Requirements and Tests: A Mobius Strip. http://www.gmelnik.com/papers/IEEE_Software_Moebius_GMelnik_RMartin.pdf
Ron Jeffries: Essential XP: Card, Conversation, Confirmation.
http://www.xprogramming.com/xpmag/expCardConversationConfirmation.htm
Tracy Reppert: Don't just break software, make software.
http://industriallogic.com/papers/storytest.pdf

Tools

Concordion: http://www.concordion.org
Cubic Test: http://www.cubictest.com
FitNesse main web site: http://fitnesse.org
FitNesse community site: http://www.fitnesse.info
Green Pepper: http://www.greenpeppersoftware.com/en/products/
JBehave: http://jbehave.org
JUseCase: http://jusecase.sourceforge.net/
PyUseCase: http://sourceforge.net/projects/pyusecase
Selenium: http://selenium.openqa.org/
StoryTestIQ: http://storytestiq.solutionsiq.com
Twist: http://studios.thoughtworks.com/twist-agile-test-automation/
Text Test: http://www.texttest.org

Mailing Lists
Agile acceptance testing (the companion mailing list to this book):
http://groups.google.com/group/agileacceptancetesting
Agile Testing: http://tech.groups.yahoo.com/group/agile-testing

243

Resources

Agile
Alliance
Functional
Testing
Tools:
http://tech.groups.yahoo.com/group/aa-ftt/
Concordion: http://tech.groups.yahoo.com/group/concordion/
FitNesse: http://tech.groups.yahoo.com/group/fitnesse/

244

Index
Symbols
.NET tools 176

A
A-7 project 54
abnormal use 51
abstract requirements 103
acceptance criteria 75
acceptance test-driven development
37
acceptance tests
and refactoring 110, 197
and user stories 169
as business resource 120
as documentation 119
as focus for development 33, 211
as specification 75, 103
as workflows 79
changing 106
code coverage 104, 235
common problems 126
communication technique 234
compared to unit tests 110, 113,
115
converting examples to 87
customer-specific 85
design hints 107
executing 115
facilitating changes 34
handover 98
incomplete 233
maintaining 126, 210
not complete verification 100

parameterisation 85
running 115, 121
saving time 209
self-explanatory 84
trust 93
what they are not 97
who should write 95
written by developers 129, 156,
235
action-oriented tests 129
affiliate advertising example 45, 59
Agile 2008 15, 25, 27, 43, 93, 183,
186, 198
agile acceptance testing
defined 31
introducing 148
not a development methodology
141
not a silver bullet 227
testing 34
Agile Alliance Functional Testing
Tools workshop 195, 199
agile development 4, 142
agile testing xxii
Amazon EC2 Cloud 189
ambiguity
manoeuvering space 212
polls 65
Andrea, Jennitta 195, 200, 201
anti-goals 71, 105
API documentation 119
architectural vision 107
Arnold, Ken 161
Asch, Solomon 20
Atlassian 178
Auer, Ken xv
Auftragstaktik 29

245

Index

automated acceptance test suites

119
automating tests 33, 86, 219
automation specialist 226

B
B-2 bomber example 21
Bache, Geoff 183, 185, 186
BandXI 178
Bay of Pigs 20
BDD (see behaviour-driven development)
Beck, Kent xxi
behaviour-driven development xxi,
38, 81
JBehave 181
user story template 160
workflow template 57
benefits
for business analysts 206
for developers 230
for testers 218
user stories 162
bet placement workflow example
55
big documents 6
black-box requirements testing xxi,
25, 97
Bluetooth example 19
brainstorming sessions 165
bridge communication strategy 205
broken window syndrome 122
Brooks, Fred 3
brushing teeth example 43
bugs
discovering 124, 223
introducing 121
user interface 152
246

verification 125
bus effect 120
business analysts
as customer proxies 62
benefits 206
challenges 209
new role 206
role name xvii
testing role 225
time 210
traditional role 206
uninterested 63
business-facing tests xxi
business goals 17, 167
business rules 32, 211
conflicts 121, 208
scattered 215
tests 132
verifying 220
business workflows 80

C
C++ tools 176
calculation tests 127, 128
card-conversation-confirmation
xxi, 164
cards for user stories 164
cargo cults 155
Carpenter, Floyd L. 22
changes
agile development 4
costs 4, 212
during development 125
facilitating 34
introducing 123, 143
propagating effects of 196
Chinese whispers 6
Clancy, Bob 11

cleaning up tests 132, 133

code coverage 104, 235
code sharing 232
cognitive diversity 20
Cohn, Mike xxii, 62, 65, 159, 213
collaboration 141, 229
Column Fixture 175, 179, 183
combinatorial explosion effect 56
commander's intent
effectiveness 30
modified template 70
common parts of tests 111
communication 133, 141, 234
bridging the gap 25
effects of problems 4
facilitating 99
gaps 5
of intent 18, 29
of requirements 3
problem 36
strategies 205
with examples 26
completeness of code 104
complexity of tests 133
Concordion xxii, 80, 179, 194
conflicts between business rules
121, 208
Confluence 178
conformity experiment 20
Connextra story card format 165
continuous integration 116, 121,
123
costs
due to misunderstandings 14
of changes 4, 212
Crain, William 30
credit card transaction example 57
Crevasse of Doom presentation 28

Crispin, Lisa xxii

Crosby, Philip 217
CRUD example 161
Cruise Control 123
CubicTest 190, 201
Cucumber 183
Cunningham, Ward xv, xxii, 54, 80,
173, 178
curly braces 161
customer acceptance testing 34, 37
customer benefits 163
customer representatives
not domain experts 62
not present during workshops
72
customer-specific acceptance tests
85
customer tests xxi

D
Dassing, Andy 113
database dependencies 130
Davenport, Tom 22
defects 124
deliverables 166
delivery plan 168
describe-demonstrate-develop xxii,
142
design
domain-driven 109
evolving 109
feature-driven 109
hints 107
improving 119
model-driven 109
detail in tests 84
developers
benefits 230
247

Index

challenges 232
collaboration 229
new role 229
perceptions of 205
repeating mistakes 218
role name xvii
suggesting new features 105
time 232
developer-testers 226
development
cooperative 222
progress 208
target 231
direct domain mapping 198
discount offers example 88
distilling specifications 80
document object model 188
DoFixture 198
DOM (see document object model)
domain
direct mapping 198
language 66, 111, 134
understanding 219, 229
domain-driven design 67, 109, 112,
198
domain knowledge
of testers 224
sharing 64
transfer 61
domain-specific languages 193
domain-specific testing languages
93, 184
virtual machine 195
dominoes example 15
DoubTech 178
DSL
(see
domain-specific
languages)
Duan, Shane 181
248

duplication in acceptance and unit

tests 114, 235

E
EC2 Cloud 189
Eclipse IDE 190, 196
plug-ins 178
edge cases 33, 47, 50, 231, 233
numerical 51
suggested by developers 49
technical 113
editor tools 199
Eisenhower, Dwight 64
epics 165
equivalence hypothesis 26
Evans, Eric xxii, 67
evolving language 134
example
affiliate advertising 45, 59
B-2 bomber 21
bet placement workflow 55
Bluetooth 19
brushing teeth 43
credit card transaction 57
CRUD 161
discount offers 88
dominoes 15
flower shop 52, 82
foreign currency exchange 14
fraud detection 49, 166
free delivery 52, 82, 84
gold-plating 12
media clips 49
poker 14
printing in Java 28
real-time reporting 30
risk profile 66
rounding to two decimals 14

search system 114

shopping carts 84
string tokenisation 114
toothbrush 43
video tour 19
VIP customers 49, 52
Windsor knot 44
example-driven development 38
examples
as acceptance tests 32
building a shared understanding
32
choosing right set 76
communicating with 26
converting to tests 82, 87
differences in similar 52
extending 131
hard to automate 90
identifying important 50
realistic 48
real-world 32
relationship with tests and
requirements 27
reviewing 72
sign-off 76
executable code 119
executable specifications 38, 113
experiment
conformity 20
five-pointed star 10
line length 20
peer pressure 20
US Army 9
exploratory testing 152, 220
extending examples 131
Extreme Programming xxii

F
facilitators 70, 152
fake objects 115
fear of blame 212
feature-driven design 109
feedback exercises 65
feedback from tests 237
ferry communication strategy 205
FIT xxii, 36, 54, 88, 96, 173, 194,
196
table 214
user interfaces 91
FIT.NET 198
FitClipse 178
FitLibrary xxii, 174, 198, 201
FitNesse xxii, 36, 96, 176, 196
user interfaces 91
version control system 200
web sites 179
five-pointed star 10
Fixture Gallery 178
fixtures 110
Concordion 180
domain 198
domain-specific languages 194
FIT 175, 179
WebTest 191
flexibility during development 125
Flowers, Jay 178
flower shop example 52, 82
foreign currency exchange example
14
Fowler, Martin 28, 109, 205
Framework for Integrated Testing
(see FIT)
fraud detection example 49, 166
free delivery example 52, 82

249

Index

functional design documents 103

functional gaps 230, 234
identifying 60, 104, 105
functional test-driven development
37
future tools 193

G
Gause, Donald xxi, 4, 9, 25, 51, 64,
65, 97
Geras, Adam 195
given-when-then 57, 81, 181
glue code 110
gold-plating example 12
good fight 154
Google groups 155
Google Tech Talks 217
Google Test Automation conference 120
Gottesdiener, Ellen 153, 206
Green Pepper 178, 200
Gross, Paul 187
groupthink 20

H
handover material 98, 212
Hatoum, Andy 145
Hendrickson, Elisabeth 199
Hewlett-Packard 22
Huggins, Jason 187
human-readable tests 87
Hunt, Andrew 122

defined xvii
relationships 221
inconsistencies 104
incremental development 143
infrastructure constraints 163
interdependent tests 131
intermittent failures 130
iterations 76, 142, 144
planning 50
specification workshops 48, 63
iterative development 143

J
Jain, Naresh 77
jargons 67
Java IDE 178
Java tools 176, 179, 181
JBehave 181
Jeffries, Ron xxi, 25, 164
JetBrains 123
JIRA 178
job security 223
JUnit 179, 182
JUseCase 185, 194
just-in-case code 33, 104

K
Kelling, George 122
Keogh, Liz 181
Kerievsky, Joshua xxi
Klein, Gary 9, 30, 65, 70

L
I
IDE test management tools 180,
182, 196, 199, 236
implementation team 97

250

lean
production process 217
software development xxii
legacy systems 185

line length experiment 20

live specification 76, 119
log files 184, 185
long tests 127

M
mailing lists 155
Mantel, Gilles 15
Marcano, Antony xxii, 226
telephone game 6
Marick, Brian xxi, 37, 151, 201
Martin, Micah 176
Martin, Robert C. 26, 176
McConnell, Steve 156
media clips example 49
Melnik, Grigori 26
mentors 154
metrics 223
Microsoft tester-developers 226
Miller, Jeremy D. 178
Minesweeper Kata screencast 186
Mission Command 29
mistake-proofing 86
misunderstandings 8
costs 14
mock objects 115, 130
model-driven design 109
Mugridge, Rick xxii, 54, 80, 120,
174, 178, 198, 201
multiple active versions 200

N
Neumann, John von 142
neutral facilitators 153
new features 105
normal use 51
North, Dan xxi, 28, 38, 181, 205

numerical edge cases 51

O
Object Mentor xxii, 176
obviousness 9
OmniGraffle 201
Ono, Taichi xxii
open issues
at workshops 72
automating tests 89
order of execution 131
organising tests 134, 200

P
Page Object 95
paralysis by analysis 160
parameterisation of acceptance tests
85
parameters with same values 128
Parnas, David 54
Pearl Harbor 20
peer pressure experiment 20
performance tests 90, 117
not automated 214
Peterson, David xxii, 80, 84, 179
phased implementation 151
Phoenix, Mickey 43, 93, 184
placeholders 169
planning poker 65
Platt, Lew 22
poker example 14
Poppendieck, Mary xxii, 105, 170,
217
Poppendieck, Tom xxii, 105
pre-planning 146
processing workflows 54, 79
product owners xvii, 15

251

Index

professional writers 97
Profit 197
progress tracking 208
project planning 165
changing the plan 169
Prussian military tactics xxii, 29
Python tools 176
PyUseCase 185, 194
Pyxis Software 178

Q
QCon 2007 28, 205
quality 217
building in 219

R
Rady, Ben 104
random values 130
realistic examples 48, 142
getting the most out of 59
specification workshop output
63
verifying 75
real-time reporting example 30
real-world examples 32
reasonable use 51
record-and-replay 99, 188
red-green-refactor 110
reducing complexity 133
refactoring 109, 134, 196
regression tests 119, 120, 123, 130
building gradually 125
don't disable 121
moving to acceptance tests 124
tools 185
relationships 221
remembering specifications 6

252

requirements
already a solution 18
black-box tests 25, 97
checking 25
documents 43, 64, 103, 210
eliciting 209, 229
most important information 25
out-of-date 119
relationship with tests and
examples 27
tests 51
traditional 103
reusing parts of tests 111
risk profile example 66
roles xvii
user interfaces 195
rounding to two decimals example
14
RSpec 183
Ruby tools 183, 201
running tested features 104
Russell, Bertrand 12

S
scalability tests 90, 117
not automated 214
scenarios
BDD 214
JBehave 182
scope of project 160, 166
Scott, Mike 50, 104, 124, 225
scripts 133
Scrum xxii
ScrumMaster as facilitator 153
SDET 226
search system example 114
security probing 117, 214
Selenese 187

Selenium 91, 94, 187

Grid 189
IDE 188
Remote Control 189
self-explanatory acceptance tests 84
set-up section 111, 132
shared understanding 44, 60, 64, 66
Shattuck, Lawrence G. 9
Shingo, Shigeo xxii, 86
shopping carts example 84
Shore, Jim xxi, 142
showstopper bugs 125
sign-off 75, 161, 212, 215
similar tests 128
Smalltalk tools 176
SMART acronym 77
smoke testing 209
software design engineer in test 226
software
developers
(see
developers)
software in production blocks 167
SolutionsIQ 184, 191, 194
source code 119
source tests 86
specification by example xxi
specifications
acceptance tests 103
compared to business goals 15
compared to scope 160
distilling 80
executable 38
flexible 126
live 119
maintaining 210
out-of-date 119
prefer to scripting 128, 133
read by developers 206
reviews 97

sign-off 75
traditional 7
specifications over scripting xxii
specification workshops 104, 150
compared to design workshops
69
domain language 68
facilitating discussion 207
facilitators 70, 152
feedback exercises 66
for user stories 170
implementation issues 64
in iterations 144
introduced 60
keeping focused 70
open issues 72
output 63
preparing for 61
reviews 72
role of customers and business
analysts 61, 64
running 60
what they are not 68
who should attend 60
speed cameras 223
spine stories 50
Spirit of Kansas, The 21
state machines 127, 128
steps in JBehave 182
Stewart, Simon 95
Stockdale, Mike 198
StoryTeller 178
storytest-driven development xxi,
37
StoryTestIQ 191
string tokenisation example 114
stub objects 115
style wars 161
253

Index

SuiteFixture 201
Surowiecki, James 20

T
tables 54
compared to natural language 57
tags 201
Talevi, Mauro 181
TDD (see test-driven development)
TeamCity 123
team leader as facilitator 153
team relationships 221
tear-down section 111, 132
technical constraints 163
technical edge cases 113
technical reviews 97
telephone game 5
test automation tools 173, 193
test cases 25
TestDriven.NET 178
test-driven design 106
test-driven development xxii, 34
test-driven requirements 15, 37
test editing tools 199
tester-developers 226
testers
analysis role 225
benefits 218
career paths 224
challenges 222
influence 222
job security 223
new role 217
role name xvii
traditional role 6
testing
agile acceptance testing 34
control 222
254

negative connotation 36, 150,

209
remote team 217
security 117
usability 117
TestNG 189
test robots 220
tests
action-oriented 129
automation 86
business-facing xxi, 37
cleaning up 132, 133
code coverage 235
complex 133
converting from examples 82
exploratory 152, 220
focused on implementation 129
human-readable 87
interdependent 131
level of detail 84
long 127
low-level 129
multiple active versions 200
not automated 117
organising 134, 200
performance 90, 117
regression 119, 120, 123
relationship with examples and
requirements 27
scalability 90, 117
separating description and automation 90, 110
similar 128
translation 88
unreliable 130
usability 90
test tags 201
TextTest 183, 200

Thomas, David 122

ThoughtWorks xviii, 123, 187, 197
three Cs of user stories 164, 171
time
business analysts 209, 210
developers 232
freeing up 220
timing issues 130
tools
current 173
future 193
IDE integration 236
Toyota Production System xxii, 86
traceability 215
training 155
transactions 130
translation layer 198
translation of tests 88
trust 93
Twist 197

U
UAT (see user acceptance testing)
ubiquitous language xxii, 66, 67,
134, 199
consistent use 111
UML 56
unanswered questions
at workshops 72
automating tests 89
understanding the domain 44, 60,
64, 66
unit tests 35, 129, 234
compared to acceptance tests
110, 113, 115
effective 116
feedback 237
purpose 114

unreasonable use 51
unreliable tests 130
usability
testing 117
usability tests 90
not automated 214
US Army xxii, 9, 29
use cases
compared to user stories 161
verifying 50
user acceptance testing 34, 37
user interfaces 214
bugs 152
tailored to roles 195
testing through 91
test robots 220
workflows 94
user interface testing 91, 227
record-and-replay 184
recorders 194
web 187
user stories xxii
and acceptance tests 169
as placeholders 169
benefits 162
business goals 167
cards 164
compared to use cases 161
easily understood 213
first cut 165
introduced 160
verifying 50
US Naval Research Lab 54

V
version control systems 135, 216
video tour example 19
VIP customers example 49, 52
255

Index

Visual Studio 178

visual workflow descriptions 201

W
Wang, Jie Tina 187
waterfall development 142
WebTest fixtures 191
Weick, Karl 70
Weinberg, Gerald xxi, 4, 9, 25, 51,
64, 65, 97
Wilson, James Q. 122
Windsor knot example 44
wireframes 91
workflows
avoiding combinatorial explosion of cases 55
BDD template 57
business 57, 80
in user interfaces 91
large 81
maintaining tests 80
processing 54, 79
testing 185
user interface 94
visual descriptions 201
with several rules 127
workshops (see specification workshops)

X
xUseCase 184
compared to domain-specific
languages 93

Y
Yahoo groups 155

256

Z
zero quality control xxii, 86
ZiBreve 178

The story continues

online
You can find more about agile acceptance testing and specification by
example on the companion web site http://www.acceptancetesting.info,
where you can also register your copy of the book to get free PDF updates.
Sign up to the Agile Acceptance Testing mailing list on Google Groups
(http://groups.google.com/group/agileacceptancetesting) to discuss this
book with other readers.
For more from the same author, see http://gojko.net.

257

258