How to use this book
Who this book is for
This book is for people who are responsible for the quality of AI output: planners who use a model to produce a report in the same format every week, marketers who refine large volumes of brand copy, researchers who need to extract evidence from long documents, maintainers of shared team prompt templates, and practitioners designing features just before API integration. They all share one need: to produce results of a consistent standard every time, rather than a result that happens to work once.
Prompt & Context Engineer (PCE) is the Yellow Belt course. It assumes White Belt (AI Literacy Foundation) fundamentals and at least three months of using generative AI in real work. You can read it without knowing how to code. Wherever API names and parameters appear, the concept is explained first and the exact wording from the official documentation follows. The aim is to let you discuss design decisions with developers in the same language, even if you are not a developer yourself. Model training and infrastructure setup are outside its scope.
After reading the book and completing the hands-on tasks, you should be able to do five things.
- Design instructions that state the task goal, constraints and success criteria.
- Make output more consistent with examples, structured tags and output formats.
- Arrange long documents, conversations and tool results within a context budget.
- Validate prompt changes with evaluation criteria and test sets.
- Diagnose and fix failure patterns such as missed instructions, broken formatting and overgeneralization.
How this book is organized
The main text has five chapters in the same order as the exam domains. Chapter 1 covers instructions, Chapter 2 examples and formatting, Chapter 3 material outside the instructions, Chapter 4 tasks that take more than one call, and Chapter 5 how to check whether a change really helped. You can read chapters separately, but they refer to one another: the success criteria in Chapter 1, for example, lead into the rubrics in Chapter 5.
Every chapter has the same ten sections.
| Section | Purpose |
|---|---|
| Learning objectives | What you should be able to do after the chapter |
| Introduction | A short story showing why the chapter matters |
| Key concepts | Explanations of the knowledge and skills in the exam scope |
| Going deeper | Trade-offs, exceptions and failure analysis |
| At work | How a fictional organization works through a problem |
| Common mistakes | Plausible but mistaken ideas |
| Summary at a glance | A table summarizing the whole chapter |
| Check the official docs | Where to read the original sources |
| Self-check | Six to eight review questions |
| Hands-on tasks | Two tasks lasting 30–60 minutes each |
Appendix A is the glossary, Appendix B contains self-check answers and explanations, Appendix C is the exam day guide, and Appendix D lists the official documents cited in this book.
The exam at a glance
| Item | Promotion exam | Free practice set |
|---|---|---|
| Questions | 30 | 10 |
| Time limit | 45 minutes | 20 minutes |
| Pass mark | Raw score of 70 or more | None |
The domain weights are as follows.
| Chapter | Domain | Weight |
|---|---|---|
| Chapter 1 | Instruction design | 25% |
| Chapter 2 | Examples and structure | 20% |
| Chapter 3 | Context engineering | 25% |
| Chapter 4 | Reasoning and multi-step tasks | 15% |
| Chapter 5 | Evaluation and improvement | 15% |
Questions use five formats. Every format is graded as a whole question, with no partial credit awarded.
| Type | Format | Grading |
|---|---|---|
| Single choice | One correct answer from four or five options | Credit only for the correct option |
| Multiple choice | Two or three answers from five or six options | Select every correct option and no incorrect ones to earn credit |
| True or false | Mark three or four statements, each combining a judgment and its reason, Yes/No | A statement is Yes only if both judgment and reason are correct; all statements must be answered correctly for credit |
| Ordering | Put four or five steps in order | The entire sequence must be correct for credit |
| Matching | Pair items on the left with options on the right | Every pair must be correct for credit; one option may remain unused |
What the questions look like
Every question opens with a concrete situation faced by a fictional company and team. One question follows the description, and conditions you must not miss are shown in bold. For example, if the condition “by changing only the instructions” is attached, solutions outside the instructions, such as post-processing or switching models, cannot be the correct answer, however effective they might be. “Without increasing latency” rules out heavy solutions that improve quality alone. More than half of the questions ask you to weigh two or more conditions against each other. They are designed so that changing a single condition in bold changes the correct answer. Rather than memorizing keywords, you need to explain why a choice is right under those conditions.
The wrong options are not random sentences. Each one carries a misconception common in real work: the belief that emphasis makes the model comply, the idea that a large context window means you can put everything in, the illusion that an answer is true because it keeps repeating, or the assumption that a higher average score means improvement. Each chapter’s “Common mistakes” section collects these misconceptions. If you can explain why an option is wrong, you can also answer new questions of the same kind. After a practice set, the review screen shows explanations and source documents for each option. Read them even for questions you got right, and check the reasoning behind your judgment.
Six-week study plan
Allow 40 minutes a day, five days a week. The hands-on tasks need only a chat interface and a spreadsheet. If you have API access, try repeating the same tasks through the API.
| Week | Reading | Hands-on work | Check |
|---|---|---|---|
| Week 1 | How to use this book; Chapter 1 key concepts | Chapter 1 task 1 | Chapter 1 self-check 1–4 |
| Week 2 | Rest of Chapter 1; Chapter 2 | Chapter 1 task 2; Chapter 2 task 1 | Finish Chapter 1 self-check; take the free practice set once |
| Week 3 | Finish Chapter 2; first part of Chapter 3 | Chapter 2 task 2; Chapter 3 task 1 | Chapter 2 self-check |
| Week 4 | Rest of Chapter 3 | Chapter 3 task 2 | Chapter 3 self-check |
| Week 5 | Chapter 4; first part of Chapter 5 | Both Chapter 4 tasks | Chapter 4 self-check; second free practice set attempt |
| Week 6 | Chapter 5; appendices; full review | Both Chapter 5 tasks | Chapter 5 self-check; Appendix C; mock exam |
Key point
Keep the output of every hands-on task in one folder or document. The tasks connect: the prompt you improve in Chapter 1 becomes the subject of evaluation in Chapter 5. At the end of six weeks, you will have your own set of prompts and a test set to validate them.
Notation and sources
- Verification date. Product features, parameters, limits and support schedules were checked again against each company’s official documentation on October 20, 2026. Appendix D groups all the pages checked by vendor.
- Principles first. AI product names and features change often. This book avoids information that dates quickly, such as model version numbers and prices, wherever possible, and focuses on design principles that remain useful for years. Statements limited to particular models explicitly say so, for example, “on some newer models.”
- Original notation. Product and API names such as
output_config.effort,reasoning.effortandmax_tokensretain their original notation. Conceptual terms are explained on first use. - Fictional organizations. Every company and institution in the work scenarios and hands-on tasks is invented. They have no connection to actual businesses or customers.
- Changed addresses. Anthropic’s current developer documentation URL is
platform.claude.com/docs, and OpenAI’s isdevelopers.openai.com/api/docs. Old addresses redirect to the new ones, but all links in this book use the current addresses.