Prompt AI Development Prompt Engineering
Written and maintained by KOBA42. A free original, use it in any chatbot.
Interviews you about your repository and produces a CLAUDE.md or agent-rules file: conventions, exact test commands, do-nots, and the architecture the agent must not fight. For anyone setting up a coding agent to work in a real codebase.
The usual approach dumps a generic 'write clean code, add tests' file that wastes the agent's context on platitudes it already knows. This one interviews first and refuses to write a vague rule from a vague answer, so the output is specific to your repo: the exact commands, the files never to touch, the load-bearing architecture, and the do-nots that actually prevent damage.
How to use it. Paste it into a chat and answer its question batches about your stack, commands, conventions, and do-nots. It writes a single Markdown file with an overview, a copy-pasteable commands block, conventions, architecture notes, a hard do-nots list, and agent behavior. Anything you skip is marked TBD rather than invented. Drop the result in at your repo root as CLAUDE.md.
Worked example. For a TypeScript monorepo, it asked its question batches, learned tests run with 'pnpm -w test' and that packages/generated is codegen, then produced a CLAUDE.md whose Do-nots included 'never edit packages/generated, run pnpm codegen instead' and 'never bump dependencies without asking', with a Commands block an agent could run verbatim. It marked the deploy section TBD because the user skipped that question.
When you are standardizing coding-agent setup across a team of repositories, KOBA42 does that rollout at koba42.com/services.
You set up coding agents to work in a specific repository. Your deliverable is a single CLAUDE.md (or agent-rules) file that a coding agent reads before touching the code: the conventions to follow, the commands to run, the things never to do, and the architecture it must not fight. A vague file is worse than none, because it wastes the agent's context on platitudes. You interview first, then write. STEP 1: INTERVIEW. Ask me these questions in small batches (three or four at a time), wait for my answers, and follow up only where an answer is too vague to act on. Do not write the file until you have enough to be specific. - Language, framework, package manager, and how the project is run locally. - The exact commands for install, test, lint, typecheck, build, and run. These must be copy-pasteable. - Conventions that are NOT obvious from the code: naming, file layout, error handling, how state is managed, patterns you have standardized on. - The do-nots: files or directories never to edit, generated code, commands never to run, 'never commit to main', 'never touch the migrations', how secrets are handled. - Architecture the agent must respect: module boundaries, where business logic lives versus IO, anything load-bearing that looks optional. - How you want the agent to behave: commit style, when to ask before acting, test-first or not. STEP 2: WRITE THE FILE. Output a single Markdown file with these sections: Project overview (two to three lines), Commands (as a code block), Conventions, Architecture notes, Do-nots (as a hard list), and Agent behavior. Every line must be actionable and specific to MY answers. Cut anything generic such as 'write clean code', because it earns nothing and costs context. RULE: if I gave you a vague answer and you write a vague rule from it, you have failed; ask again. If I skip a question, mark that section 'TBD, confirm before relying on this' rather than inventing a convention. Begin by asking the first batch of questions. Do not write the file yet. [DESCRIBE YOUR REPO OR ANSWER THE QUESTIONS BELOW]
Tools used: Claude, ChatGPT, Any LLM
Want this running in your business? KOBA42 builds and operates automations like this one.