How to use this book
Who this book is for
AI App Builder (ABD) is the Blue Belt course for people who delegate tasks to a coding agent to build a small web app and take responsibility for the result: running it, verifying it, fixing it and deploying it. This book is written for three groups of readers.
- Planners and founders building their first service with a coding agent
- People without a computer science background running a side project after work or at weekends
- Junior developers starting to work alongside agents
You do not need much preparation. White Belt (AI Literacy Foundation)–level understanding, the ability to type commands in a terminal and basic browser skills are enough. You do not need to know programming language syntax. This course develops your ability to judge whether delegated work is correct and undo it when it is wrong, rather than your ability to write every line of code yourself.
A coding agent is an AI tool that, once you describe a task in plain language, reads files, edits code and runs commands on its own until the job is done. OpenAI's Codex, Anthropic's Claude Code and Google's Gemini CLI are representative examples. Their screens and command names differ, but the judgments a person must make are much the same: define what counts as done, check results against evidence, record changes, restrict dangerous permissions and remain responsible after putting the app in front of users.
This book therefore teaches decision principles that work in any tool before explaining how to use specific UI controls. Tool-specific features are examples of putting those principles into practice. Large-scale distributed system design and model training are outside its scope.
Note
A coding agent is like a new colleague who works fast and never gets tired. They work diligently, but unless you tell them, they do not know what "done" looks like or how far they are allowed to go. The five chapters show you how to assign that colleague a clearly defined task (Chapter 1), check the result (Chapter 2), keep a record (Chapter 3), block risks (Chapter 4) and put the work in front of customers (Chapter 5).
How this book is organized
Chapters 1 to 5 follow the same order as the exam domains, and each chapter is organized into the same ten sections. Each is written so that you can understand it on its own, whichever chapter you begin with.
Each chapter contains learning objectives, an introduction, key concepts, going deeper (edge cases and failure analysis), at work (decisions in invented organizations, including rejected options), common mistakes, a summary at a glance, official documentation to check, self-checks (explained in Appendix B) and hands-on tasks (30–60 minutes), in that order.
Appendix A is a glossary; Appendix B explains the self-check answers; Appendix C is an exam day guide; Appendix D lists the official documents cited in this book.
The exam at a glance
The ABD course offers three modes.
| Mode | Questions | Time | Features |
|---|---|---|---|
| Practice | 10 questions | 20 minutes | Practice for free on the web and read an explanation for each question |
| Mock exam | 30 questions | 50 minutes | Practice in the app under conditions like the real exam |
| Promotion exam | 30 questions | 50 minutes | Take the exam in the app and earn the Blue Belt if you pass |
The pass mark is a raw score of 70 or more. Questions are worth 2 or 3 points according to difficulty, displayed beside them as [2 points] or [3 points]. Your raw score is your earned points scaled to a score out of 100, so missing a 3-point question costs more than missing a 2-point one.
The domain weights are as follows.
| Chapter | Domain | Weight |
|---|---|---|
| Chapter 1 | Defining tasks and instructing agents | 25% |
| Chapter 2 | The run-and-verify loop | 25% |
| Chapter 3 | Version control and change review | 20% |
| Chapter 4 | Security and dependencies | 15% |
| Chapter 5 | Deployment and operations | 15% |
Caution
In the promotion exam, you must answer questions in order. Once you confirm an answer, you cannot return to that question. The mock exam lets you move freely between questions, so develop the habit of treating confirmation as final even while practicing. Appendix C gives detailed advice on allocating your time.
What the questions look like
Every question is answered by choosing one of ①–⑤. There are five question formats.
| Format | What it looks like | How to approach it |
|---|---|---|
| Single choice | Choose the most appropriate of five options | Find the option that satisfies all the conditions |
| Combination choice | Choose the combination containing all appropriate items from a–e in the statements | Judge each item separately, then select the combination |
| Statement evaluation | Decide whether both the judgment and its reason are correct for each statement | A statement is wrong if the reason is wrong, even when the conclusion is right |
| Ordering | Put steps (a)–(e) in the correct order | Ask what each step requires beforehand |
| Matching | Connect the items on the left with those on the right | Eliminate definite pairs first. One item on the right is often left over |
The conditions in bold in the question are the key to choosing the correct answer. Change a bold condition such as "first," "while preserving the record" or "protect locations outside the working folder," and the answer changes too. The following ordering example was written for this book.
You want to turn a folder containing only `README.md` into a repository and push it to a remote for the first time.
Team rules require registering the remote address after making the first commit.
Which sequence puts the steps in the correct order? [2 points]
(a) git commit -m "Initial commit"
(b) git init -b main
(c) git push -u origin main
(d) git add .
(e) git remote add origin <remote-address>
Ask "What must already exist for this step?" in turn. You need a repository before you can stage files (b→d), staged content before you can commit (d→a), and a registered remote address before you can push (e→c). Normally (e) could happen any time after (b), but the condition on the second line places it after the commit. The answer is therefore (b)-(d)-(a)-(e)-(c). Real questions similarly use explicit conditions to determine a unique order, so read every condition carefully.
Six-week study plan
This plan assumes 40 minutes a day, five days a week. Do the hands-on work as well as the reading. Most questions in this course describe situations that are easier to judge if you have experienced them yourself.
| Week | Reading | Hands-on work |
|---|---|---|
| Week 1 | How to use this book, Chapter 1 key concepts | Create a practice repository and write an instruction file |
| Week 2 | Rest of Chapter 1, beginning of Chapter 2 | Delegate a task with a definition of done; run locally |
| Week 3 | Rest of Chapter 2 | Fix a deliberately planted bug using failing output |
| Week 4 | Chapter 3 | Review diffs, try three ways of undoing changes, recover with reflog |
| Week 5 | Chapter 4, one practice set | Check secrets, set permission rules and a sandbox, write your first review notes |
| Week 6 | Chapter 5, appendices | Deploy a preview, roll back and promote, take a mock exam and review mistakes |
Key point
The self-checks were written separately to avoid overlapping with exam questions. Rather than memorizing answers, learn the "why" in the Appendix B explanations. The exam tests your ability to apply principles to unfamiliar situations.
Notation and sources
- Product and feature names stay in their original form (for example, plan mode, Instant Rollback,
AGENTS.md). Commands and file names usemonospace. - Tool-specific features, defaults and plan requirements are included only where confirmed in the vendors' official documentation on October 20, 2026. The pages checked appear in each chapter's "Check the official docs" section and in the complete list in Appendix D.
- Coding agents and deployment platforms change quickly. If a menu name or default differs from this book, the official documentation takes precedence. The principles change less often, so the reasons for following them remain useful even when features change.
- All organizations, apps and people in the examples are invented. They have no connection with real companies or individuals.
- All example keys and passwords are fake. Use test values rather than real secrets in the exercises too.