Copy-Paste CLAUDE.md Templates for 8 Project Types

A CLAUDE.md template is the instruction file you drop in the root of your project so Claude Code knows how your project works before it touches a single line. Claude Code reads CLAUDE.md at the start of every session. That makes it the memory and the house rules for the agent, and the single highest-leverage file you will ever write for a build.

I treat it like the brief I would hand a new contractor on day one. Without it, the agent guesses your stack, your naming, and your commands, and it guesses differently every session. With it, the agent shows up already knowing the project. The same prompt gets you a far better result, and you stop repeating yourself.

Every strong CLAUDE.md follows the same five-part shape. We call it the CCC project-brief skeleton, and once you can see the five parts you can write one for any project in a couple of minutes. Each part answers a question the agent would otherwise have to guess.

Here are the first four CLAUDE.md templates. Each follows the CCC project-brief skeleton. Open Claude Code in the desktop app, create a CLAUDE.md in your project root, paste the one that fits, and swap the generic details for your real ones.

1) Web app: '# Project. A web app for [users] to [job]. Stack: React + TypeScript on the front end, Node on the back end, Postgres for data. Conventions: function components, hooks over classes, colocate tests next to source. Commands: npm install, npm run dev, npm test, npm run build. Do: keep components small and typed. Don't: add a new dependency without flagging it; never commit secrets.'

2) REST/GraphQL API: '# Project. A REST API serving [clients]. Stack: Node + Express, Postgres via an ORM, JWT auth. Conventions: one route file per resource, validate every input, return typed error shapes. Commands: npm install, npm run dev, npm test, npm run migrate. Do: write a test for every new endpoint. Don't: log secrets or tokens; never ship an endpoint without auth and rate limiting.'

3) Data/ML project: '# Project. A pipeline that [trains/scores] on [data]. Stack: Python, pandas, scikit-learn, notebooks for exploration and scripts for production. Conventions: raw data is read-only, every transform is a pure function, fix the random seed. Commands: pip install -r requirements.txt, python train.py, pytest. Do: log metrics for every run. Don't: edit files in data/raw; never commit datasets or model weights.'

4) Mobile app: '# Project. A mobile app for [platform] that [job]. Stack: React Native + TypeScript, REST API back end. Conventions: screens in /screens, shared UI in /components, no business logic in views. Commands: npm install, npm run ios, npm run android, npm test. Do: handle offline and loading states. Don't: hardcode API URLs; never block the main thread.'

Here are the last four CLAUDE.md templates, covering the project types most builders and agencies actually ship. Same skeleton, same paste-and-edit move.

5) Marketing/landing site: '# Project. A landing site for [product] built to convert. Stack: static site generator, plain CSS or Tailwind, deployed to a CDN. Conventions: one section per component, copy lives in content files, keep it fast. Commands: npm install, npm run dev, npm run build. Do: keep Lighthouse green and images optimized. Don't: add heavy JS for effects; never break the mobile layout.'

6) Automation + scripts: '# Project. A set of scripts that automate [task]. Stack: Python or Node, run on a schedule. Conventions: one script does one job, config in env vars, every script is idempotent. Commands: pip install -r requirements.txt, python run.py. Do: log what ran and what changed; make every run safe to repeat. Don't: hardcode credentials; never delete data without a dry-run flag.'

7) Open-source library: '# Project. A library that [does X] for other developers. Stack: TypeScript, zero runtime dependencies where possible. Conventions: semantic versioning, public API documented, breaking changes flagged. Commands: npm install, npm test, npm run build, npm run lint. Do: write a test for every public function and update the changelog. Don't: change the public API without a major version bump.'

8) Agency client repo: '# Project. Client work for [client]. Stack: [their stack, not ours]. Conventions: match the client's existing style exactly, branch per task, descriptive commits. Commands: [their install/run/test/deploy]. Do: ask before introducing a new tool; keep the handoff clean and documented. Don't: refactor outside the ticket; never push to main or deploy without sign-off.'

CLAUDE.md earns its keep because of how Claude Code runs. The agent works in a loop - it gathers context, acts, then verifies. CLAUDE.md is the context it gathers first, every single time, and the Commands section gives it the way to verify its own work. A good CLAUDE.md makes both ends of that loop sharper.

If you want to see that cycle in detail, read [what are agent loops in Claude Code](/blog/claude-code-agent-loops). For the broader habits that separate fast builders from frustrated ones, [how to use Claude Code like a pro](/blog/how-to-use-claude-code-like-a-pro) is the next step.

A CLAUDE.md is a living file, not a one-time setup. The fastest way to rot one is to write it on day one and never touch it again. Treat it like the project's running memory and it keeps paying off.

Skipping the file entirely is one of the most common beginner errors, and it is covered in [Claude Code mistakes beginners make](/blog/claude-code-mistakes-beginners-make). A stale or missing CLAUDE.md is why two people get wildly different results from the same prompt.