This file is runtime-mode guidance for Claude Code. It's what agents (subagents in .claude/agents/, skills in .claude/skills/) and runtime Claude Code sessions see. It's committed and shared with anyone who clones the repo.
Builder-mode context — the agent-building playbook, LikeC4 structural model imports, decision history — lives in
CLAUDE.local.md. That file is gitignored. Run./bootstrap.shto create it from the example.
Rules an agent has to honour while doing work. Violating one means an agent stepped outside designed behavior.
- When docs disagree, follow docs/AUTHORITY-ORDER.md.
Add runtime-relevant project invariants here as they emerge. Each needs a why so future sessions can judge edge cases. Example shape: "All email sends go through
bin/send.shso retries are centralised. Why: prior incident where two skills retried independently and double-sent."
Filesystem conventions agents and CLIs depend on. Create these directories when you add your first agent or skill.
.claude/agents/— repo-defined subagent prompts.claude/skills/— skills (canonical surface and runtime).claude/context/— runtime reference docs that prompts load inlinebin/orscripts/— CLI tools (deterministic, no LLM call)
At runtime, invoke skills directly via Skill(...) instead of searching for SKILL.md files.
Replace this section with project-specific runtime context once you've forked the template: a one-paragraph description of what this system does, the key entities agents work with, and the external integrations they call.
After cloning:
- Run
./bootstrap.sh— pulls submodules, installs npm deps, creates your localCLAUDE.local.mdfrom the example - Replace this section with runtime-facing project description
- Edit
CLAUDE.local.mdfor your builder context (it auto-imports the playbook) - Read QUICKSTART.md for the 7-step daily workflow
See README.md for the full template usage guide.