Want to write better AI prompts but find the tips online scattered and hard to start with? Anthropic, the company behind Claude, has released a free interactive tutorial that has passed 37,000 stars on GitHub, one of the most popular resources of its kind. It is not a pile of abstract theory but nine chapters, each with a lesson and exercises, so you learn by trying and build prompting skill from zero. Crucially, it is official, so it demonstrates exactly what "a good prompt" looks like in the eyes of the model's own makers. This article breaks down the tutorial chapter by chapter: its structure, what you learn, the examples, and how to make the most of it.
What is this tutorial?
It is Anthropic's official "Interactive Prompt Engineering Tutorial," designed to teach you, step by step and systematically, how to write optimal prompts on Claude. It has nine chapters plus an advanced appendix, each with a lesson and exercises, and a "Playground" where you can edit prompts on the spot and see Claude respond, so you feel first-hand how "changing one word" changes the output. There is also a Google Sheets version, friendlier for non-engineering backgrounds. The tutorial uses the smallest, fastest and cheapest model in the Claude family, so you can practise repeatedly without burning money.
Why it's worth learning: from "copying prompts" to "building prompts"
Plenty of "god-tier prompt templates" circulate online, but copy one and it stops working the moment the scenario changes, because you only have the result, not the principle. The value of this tutorial is teaching the reasoning behind a prompt: which parts it is made of, what problem each part solves, and what the common failure modes are. Master that and you stop depending on other people's templates, able to build from scratch for your own tasks. For a business, this skill transfers directly to everyday flows like customer service, reporting and document processing, a far higher return than memorising a few templates.
Chapters 1 to 3: build the foundation
The first three chapters teach the fundamentals. Chapter one covers "basic structure," how a good prompt should be organised: where the role, task, context and output requirements each go. Chapter two teaches being "clear and direct," for example not writing "take a look for me" but "identify three clauses in the contract below that disadvantage the buyer," reducing ambiguity. Chapter three teaches "assigning roles," such as having the AI act as a senior legal counsel or financial analyst so the tone and depth fit the context. Basic as they are, these are exactly the parts most people skip yet that most affect results.
Chapters 4 to 5: separate data from instructions, control format
The middle chapters fix the most common confusion. Chapter four teaches separating "data" from "instructions," for example wrapping the content to process in XML tags or clear delimiters so the AI does not execute a pasted customer email as if it were a command, which is also the basis for defending against malicious content (prompt injection). Chapter five teaches controlling output format, such as requiring a table, JSON or fixed sections, even "speaking for the AI" to steer it into the style you want. These two moves matter especially for automation and feeding AI output into systems.
Chapters 6 to 7: think step by step and teach by example
Chapter six teaches "precognition," asking the AI to think step by step. On complex reasoning, calculation or multi-step judgment, letting the model write out its thinking before concluding noticeably improves accuracy, because it does not rush to an answer. Chapter seven teaches "using examples," or few-shot: give a few "input-to-output" demonstrations and the AI learns the style, format and tone you want, often more effective than lengthy explanation. In practice, one good example beats ten adjectives. These two chapters are the key to going from "able to use it" to "using it accurately."
Chapter 8: avoiding hallucinations
Chapter eight focuses on reducing hallucination, where the AI makes things up or answers wrong yet confidently. It demonstrates several methods: explicit instructions, requiring cited evidence, allowing the AI to say "I don't know" when unsure rather than fabricating, and telling it to find evidence before answering. For a business, this chapter directly affects whether you dare rely on the AI's output for decisions, and it matters even more with numbers, names, and legal or medical information.
Chapter 9: building real business complex prompts
The final chapter combines all the earlier techniques to build real business-grade complex prompts from scratch, covering a customer-service chatbot and legal services, with exercises for finance and coding. This chapter is the most hands-on: you see how a complete, usable prompt is built up layer by layer, from role setup, data injection and output format to error handling and edge cases. After it, you have several real templates you can adapt and use immediately.
Advanced appendix: chaining, tools, retrieval
The appendix, "Beyond Standard Prompting," introduces three advanced capabilities. Prompt chaining: split a big task into a relay of prompts, for example summarise, then classify, then draft a reply, each step more stable. Tool use: let the AI call external tools such as a calculator or an API for things it is not good at. Search and retrieval: let the AI draw on external or up-to-date data before answering. This part is the bridge from "writing one good prompt" to "building an AI workflow," ideal for those who have the basics and want to go deeper.
How to make the most of it
Work through the chapters in order without skipping, since later chapters build on earlier concepts. Do the exercises and try editing prompts by hand in the Playground, feeling how the output shifts after a one- or two-word change. Practise on your company's real scenarios, for example your own customer enquiries or reporting needs, so what you learn is immediately usable. You can also turn each chapter's key points into your own "prompt checklist" to run through whenever you write a prompt.
Common pitfalls
The three most common pitfalls: first, copying templates without learning the principle, so they fail when the scenario changes; second, cramming too many requirements into one prompt so the AI drops some, when you should split the steps or chain; third, giving no examples and no clear format, then expecting the AI to guess your intent. Avoid these three, pair them with the tutorial's framework, and you will reliably write dependable prompts. Want the tutorial link and my summarised key notes? Visit ai.ud.hk to explore UD's AI Staff solutions and see how to bring strong prompting skill into your team's workflow.
懂AI，更懂你｜UD相伴，AI不冷
