distill（设计提炼）

description: Rewrite repo CLAUDE.md into the best onboarding guide for this project

DISTILL

Distill this repo's knowledge into a single, sharp CLAUDE.md.

Purpose

• Rewrite the current repository's CLAUDE.md into the best possible onboarding + operating guide for this project.
• Output must let a new contributor start useful work in minutes, without asking questions.
• Scope is this repo only; global philosophy lives in ~/.claude/CLAUDE.md.

The Kill Test

Every line in CLAUDE.md must survive three questions:

1. Discoverable? Would Claude figure this out from ls, cat package.json, or reading code? → DELETE
2. Surprising? Does this describe something counter-intuitive or unexpected? If not → DELETE
3. Earned by pain? Did this instruction come from something going wrong? → KEEP

Start with nothing. Wait for Claude to do something wrong. Write THAT down.

What to kill

• File trees (Claude can ls)
• "This project uses React with TypeScript" (Claude can see tsconfig.json)
• Architecture descriptions that restate what code shows
• Technology stack versions (in package.json)
• Module lists (in imports)
• Development philosophy restating linter/type config
• Sprint plans, roadmaps, marketing copy, performance metrics
• Build system descriptions (in build config)
• Testing framework choice (in test config)

What survives

• "Run yarn test:unit not npm test" — weird, not discoverable
• "Don't touch src/legacy/ — three enterprise clients depend on it" — earned by pain
• "The auth middleware is load-bearing, all of it, don't be a hero" — counter-intuitive
• "DATABASE_URL not POSTGRES_URL — Prisma's Rust engine reads it before Node" — painful gotcha
• "e2e tests can't run headless due to extension limitations" — surprising constraint

Inputs

• Always read, if present:

- Repo AGENTS.md. - Existing repo CLAUDE.md. - Root README and any high-signal docs in docs/. - Key config / entry files (e.g. package.json, pyproject.toml, docker-compose.yml, main entrypoints).

Target Shape for CLAUDE.md

# CLAUDE

## Purpose
- What this repo is and why it exists.

## Architecture Map
- Main domains / services and how they fit.
- Where to start reading code (file:line).

## Run & Test
- Commands to run app, tests, lint, typecheck.
- Required env, secrets, or services.

## Quality & Pitfalls
- Definition of done, PR expectations, review norms.
- Non-obvious invariants, footguns, "never do X".

## References
- Key docs / ADRs / diagrams (paths).
- External systems / dashboards / runbooks.

The Codex First-Draft Pattern

Codex drafts the new CLAUDE.md. You review and refine.

codex exec "DISTILL: Rewrite CLAUDE.md for this repo. Read README, ARCHITECTURE.md, docs/, and key code. Follow target shape below. Output to /tmp/claude-draft.md" \
  --output-last-message /tmp/codex-distill.md 2>/dev/null

Review Codex's draft. Refine for accuracy and compression.

Algorithm

1. Gather (Codex does this)

- Read inputs; understand what the repo does, how it runs, and how it fits into the wider system.

1. Kill Test every line

- For each line, ask three questions: - Discoverable? Would Claude figure this out from ls, cat package.json, or reading code? → DELETE - Surprising? Does this describe something counter-intuitive? If behavior is what you'd expect → DELETE - Earned by pain? Did this come from something going wrong? → KEEP - Also classify: - General / global → belongs in ~/.claude/CLAUDE.md, do not restate here. - Marketing copy / roadmaps / sprint plans → DROP (not instructions)

1. Draft new CLAUDE.md (Codex produces first draft)

- Fill the target shape above with tight, repo-specific bullets. - Prefer bullets over paragraphs; every line must earn its place. - Link to existing docs instead of duplicating them.

1. Compress

- Apply the 3-2-1 test: - 3 key decisions or invariants newcomers must know. - 2 critical insights about architecture or workflow. - 1 clear starting point in the codebase. - Rewrite soft prose into sharp, actionable lines.

1. Validate

- Run the checklist below. - Only propose the new CLAUDE.md once all checks pass.

Checklist (must pass)

• Readable in ≤3 minutes; roughly ≤80 lines.
• Every line fails the Kill Test (not discoverable from code/config/ls).
• Run commands included (only if non-obvious — skip if standard pnpm dev/test/build).
• Captures footguns, gotchas, and "don't touch X" warnings.
• No file trees (Claude can ls).
• No technology stack listings (Claude can read package.json).
• No architecture descriptions that restate what code shows.
• No marketing copy, roadmaps, or sprint plans.
• Links to deeper docs / ADRs instead of copying them.
• Does not restate global behavior or philosophy from ~/.claude/CLAUDE.md.

Compression Examples

Kill (discoverable): "The project uses Turborepo with pnpm workspaces, Next.js 15, React 19, Tailwind CSS, and TypeScript strict mode." → Claude can read package.json and tsconfig.json.

Kill (not surprising): "Content scripts inject into web pages. Background scripts handle lifecycle. Core modules have specific responsibilities." → This is what every browser extension does.

Kill (file tree):

├── apps/web/       # Next.js app
├── packages/common/ # Shared types
└── turbo.json      # Turborepo config

→ Claude can ls.

Keep (earned by pain): "DATABASE_URL not POSTGRES_URL — Prisma's Rust engine reads it before Node.js starts. Runtime env modifications are too late in serverless."

Keep (surprising): "e2e tests can't run headless due to browser extension limitations."

Keep (load-bearing): "Don't touch src/legacy/ — three enterprise clients depend on it, and there are no tests."

Philosophy

• Maximum signal per word.
• Document what is not obvious from code or README.
• If you can't find 3 decisions, 2 insights, and 1 starting point, read more code before writing.

Graduation Rules

CLAUDE.md is a staging area for learnings. When sections grow, graduate them:

• Patterns section > 10 lines → extract to a skill
• Rules section > 5 items → extract to an agent
• Workflow section > 3 steps → extract to a command

Keep CLAUDE.md lean. Knowledge graduates to executable artifacts.