Claude Playbook

Claude Playbook

A working .claude/ directory for Claude Code, and
the reasoning behind every file in it.

Rules that load only when they apply, skills for review and verification,
subagents that keep exploration out of your context, and hooks that enforce what
instructions can only request.
Independent community project. "Claude" and "Claude Code" are products of
Anthropic. Not affiliated with, endorsed by, or
sponsored by Anthropic.

Contents

- Why this exists
- Before you install
- Install
- Your first fifteen minutes
- What you get
- How the pieces fit together
- The components, one at a time
  - 1. CLAUDE.md
  - 2. Rules
  - 3. Skills
  - 4. Subagents
  - 5. Hooks
  - 6. Workflows
  - 7. Settings and permissions
  - 8. MCP servers
  - 9. CI and automation
  - 10. Evals
- Make it yours
- Check your setup
- Guides
- Compatibility
- Contributing

Why this exists

Claude Code writes good code and still drifts, because nothing in your
repository tells it what "good" means here. Left alone it will patch symptoms
rather than causes, build a component that already exists three directories
over, leave dead code behind, and report that tests pass without running them.

None of that is a model failure. It is missing context, in the specific places
Claude Code gives you to put context.

This repository fills those places, and explains each choice, so you can keep
what fits your project and delete the rest.

New in 2.0 (September 2026): the hook registration format, the memory
precedence model, and half the frontmatter documented here changed since the
first release. Rules are now path-scoped, which cut the always-loaded context
by 82%. See what's new and the
changelog.

Before you install

Two things, both quick.

1. Install jq. Every hook needs it to read its…