How the Builders Prompt
Updated
IntermediateWhat you'll learn
~12 min- Locate the source of these techniques and why they are worth trusting
- Separate what you already do well from the techniques that fill a real gap
- Choose which lessons in this module to take and in what order
Module 10 taught you the foundation of prompting: be specific, give context, set constraints, describe what “done” looks like, iterate one change at a time. If you have been prompting daily, you already do most of that without thinking.
This module is the next layer — a small set of techniques that the people who built Claude reach for when a plain instruction is not getting the result. They are not tricks. Each one changes how the model processes your request, and each one earns its place by fixing a specific failure mode: vague reasoning, wrong format, confident-but-wrong answers, lost details in a long document.
This module is drawn from Anthropic’s official prompt engineering material — the techniques the people who built Claude document and teach. It exists in three free forms:
- The docs — the maintained reference (Prompt engineering overview)
- The interactive tutorial: 9 hands-on chapters with an answer key (github.com/anthropics/prompt-eng-interactive-tutorial). It was written for Claude 3 models, so where it and the docs differ, follow the docs.
- The workshop: Anthropic’s prompt engineering talk with Zack Witten (Building with Anthropic Claude, via Class Central). The docs are the maintained version; where the talk differs, follow the docs.
These techniques were written for Claude, but they generalize: role prompting, examples, and quote-first grounding improve results on every frontier model. Where a technique is Claude-specific, this module says so.
What you already have vs. what is new
Be honest about the gap. The workshop covers about a dozen techniques. Roughly half of them you already learned in Modules 1 and 10:
| Already covered (Modules 1, 10) | New in this module |
|---|---|
| Be clear, specific, and direct | Give the model a role (“You are a senior…”) |
| Provide context | Teach with examples (multishot, done well) |
| Set constraints | Set the model’s thinking effort, and ask for a decision record you can check |
| Specify output format | Structure prompts with tags, and lock the output format |
| Describe the expected outcome | Ground answers in quotes to stop hallucination |
| Iterate one change at a time | Chain prompts so each step does one job |
| Test prompts instead of trusting the first one |
You do not need to relearn the left column. This module is the right column.
Self-assessment: where is your gap?
Work through this list honestly. For each item you cannot do confidently today, there is a lesson that teaches it.
1. Prime the model’s behavior before you give it the task.
Opening a prompt with “You are a senior security reviewer auditing this code for a government deployment” changes the depth, vocabulary, and caution of everything that follows. If you only ever describe the task and never the role, you are leaving the easiest win on the table.
2. Show the model what “good” looks like instead of describing it.
For anything with a consistent shape (a report format, a commit message style, a classification), a few good examples teach the pattern better than a paragraph of instructions ever will. Most people under-use examples or write inconsistent ones.
3. Set how hard the model thinks, and ask for something you can check.
Current models already reason before they answer, at a depth you set. On anything with logic (math, analysis, multi-step decisions), raise the effort setting your tool exposes (/effort in Claude Code and Antigravity, for example), and ask for a decision record (assumptions, options, chosen path, what would change the call) rather than a transcript of the thinking.
4. Separate your instructions from your data so the model never confuses the two.
When you paste a long document plus a question, the model can blur which part is the task and which part is the input. Wrapping each part in clearly labeled tags fixes that — and lets you control the exact shape of the output.
5. Get answers grounded in the source instead of plausible-sounding inventions.
The single highest-value technique for research, legal, compliance, and analysis work: make the model pull the relevant quotes from the source first, then answer using only those quotes. It is how you turn “sounds right” into “is right.”
6. Break a hard prompt into a chain where each step does one job.
When one prompt is trying to extract, analyze, and write all at once, quality drops. Chaining — feeding the output of one focused prompt into the next — beats a single overloaded prompt on complex work.
7. Know a prompt is good because you tested it — not because the first run looked fine.
A prompt that worked once can fail on the next input. Defining what a good answer looks like and running the prompt against a handful of cases is the difference between a prompt you hope works and one you know works.
How these stack
The techniques are independent — take them in any order based on your self-assessment. But they compound. A mature prompt for a hard task often uses several at once:
[ROLE] You are a senior grants analyst.[EXAMPLES] Here are two examples of the summary format I want…[DATA] <proposal> …the full text… </proposal>[THINK] Before the summary, give a short decision record: assumptions, the risks you weighed, and what would change your read.[GROUND] Quote the exact budget lines before you total them.[FORMAT] Then output the summary as the example format above.That is one prompt using role, examples, structure, thinking, and grounding together. You do not start there: you start with one technique that fixes your current problem, and you add the others as the task demands. Use structured tag blocks like this for one-shot document analysis, grading rubrics, or reusable skills and AGENTS.md instructions; for interactive software builds, switch to Plan mode inside your session and give a short plain-English goal.
Flags, config, and exact slash-commands change between Claude Code, Antigravity, Codex CLI and Copilot CLI, and between versions. The patterns in this module do not. Learn current syntax from your tool’s own docs; learn what to reach for, and why, here.
Which of these is genuinely NEW in this module, versus something Module 10 already covered?
Key takeaways
- These are sourced techniques, not folklore. They come from Anthropic’s prompt engineering docs, tutorial and workshop. The docs are the maintained version, updated as models change; where an older source disagrees, follow the docs.
- Half of prompting you already know. Be specific, give context, set constraints, describe the outcome. This module is the other half.
- Seven techniques fill the real gaps: role, examples, thinking, structure, grounding, chaining, and testing.
- Take them in any order. Use the self-assessment to find your gap and start there.
- The biggest single win for serious work is grounding — making the model quote its source before it answers. If you read one lesson, read that one.
Next: Give the Model a Role