Extending

Three rules hold the design together.

1. An agent is a thin trigger. It says who and when, nothing more. It stays short, because its description is loaded into every session.
2. A skill is the single source of truth. The actual method lives there once, and is never copied into an agent.
3. A rule that matters becomes a gate. Enforcement sits at the tool level — a hook, a permission, a test case. The model is not asked to remember it.

When you add an agent or a skill, follow the AGENT_TEMPLATE.md contract: frontmatter (name · description · least-privilege tools · model tier), the Trigger phrases: line in the body, and the body sections (When → Expertise stance → How → Coordination → Definition of Done → Output → Escalation → Example → Constraints). smoke-test.sh refuses a component that nothing routes to, so nothing ships asleep. /crew-skill walks you through it and ends in the gates.

To take a single component rather than the full install, npx crewforth add <name> copies one agent or skill (and the skills an agent uses) into ./.claude; npx crewforth add --list shows the catalogue.