claude_code_playbook is my opinionated way of running software projects with Claude Code.
It comes in two parts:
- spec-driven-dev is a skill that settles what to build: SPEC, ROADMAP, CLAUDE.md, decisions and a log, before any code is written.
- playbook is a plugin that runs how the work happens: a per-project profile, a fixed task routine (
/task→/ship→/done), git hooks, CI, and milestone reviews.
The main decisions behind it:
- Spec first, one task per conversation, every task with its acceptance criteria and the test that proves it.
- Red/green TDD: the failing test is committed on its own before the code.
- Review by fresh eyes: three read-only reviewers in parallel, plus a security pass on sensitive code.
- The main branch only takes PRs, and nothing merges before CI is read.
- The autopilot is experimental and opt-in. It works through approved ROADMAP tasks while I’m away, one fresh session per task in its own worktree. It never merges anything on its own.
The playbook owes a lot to Addy Osmani’s blog on working with AI coding agents. Its core comes from three ideas found there: the spec comes before the code, each task stays small enough for a test to prove it, and human review is the step that actually sets the pace.
It’s MIT-licensed, experimental, and comes with no warranty: read what it does before you trust it.