CLAUDE.md, AGENTS.md, and Meta-Prompting — Closing the Series

The problem: The agent doesn’t know your project

Every AI agent starts a session from zero. It doesn’t know your stack, how to run tests, or what conventions you follow. Every time you explain the same things.

Solution: instruction files in the repository — the agent reads them automatically.

Files overview

File	Agent	Description
CLAUDE.md	Claude Code	Project memory, read automatically
AGENTS.md	Codex, Copilot, Cursor	Open standard (Linux Foundation)
DEVIN.md	Devin	Per-repo configuration
docs/	All	Detailed documentation on-demand

What should it contain?

Think of it as onboarding a new team member with amnesia:

• Commands — how to run the dev server, tests, linter
• Architecture — directories, stack, dependencies
• Conventions — commit format, code style, branching
• Important rules — what NOT to do (e.g., modifying migrations)
• Test workflow — pattern, reporting, tagging

Progressive Disclosure

CLAUDE.md should be under 300 lines. Details go in docs/:

CLAUDE.md              ← Short briefing
docs/
  architecture.md      ← Architecture details
  test-strategy.md     ← Test strategy
  api-spec.md          ← API specification

What NOT to include

• Code style rules — that’s what the linter is for
• Obvious things — “this is a React project” (visible from package.json)
• Secrets — passwords, tokens, API keys
• Entire documentation — use docs/

---

Meta-prompting: a prompt that creates a prompt

The most important takeaway from the entire series:

If you don’t know how to write a good prompt — ask AI to write it for you.

Universal template

I want to use the [FRAMEWORK] framework for [GOAL].
Context: [brief description of the situation]
Generate a ready-to-copy prompt for me.

Cheat sheet: one prompt to rule them all

I am a QA tester. I want to achieve: [GOAL].
My context: [2-3 sentences].

Suggest:
1. Which prompt framework would be the best fit
2. Which AI model/tool to use
3. A ready-to-copy prompt

This one prompt replaces all the knowledge from this series — AI will select the framework and model on its own.

---

Series summary

• A prompt framework is not magic — it’s a thinking structure
• Auto QA ≈ Dev — auto QA tools are developer tools
• Multi-agent is the future — no single agent is the best at everything
• Experiment — the same prompt in ChatGPT, Claude, and Gemini will produce different results
• Let AI write prompts for you — over time you’ll start doing it intuitively

Thank you for this journey through 16 posts. I hope at least one framework or scenario will find its way into your daily work.

Grzegorz Holak — Test Architect & AI Ambassador