Skill registry

The skill registry is the agent-config building block for repeatable procedures: reusable instruction bundles with tool and knowledge grants. The registry surface is /ai/skills. Each row carries name, key, tools, knowledge, agents, and status; the agents column shows which agents adopted the skill. Skills shape behavior; they do not grant capability. A skill built on broken tools teaches the agent to fail confidently, so the sequencing rule is load-bearing: tools first, skills second, wiring verified last.

• Each skill encodes one procedure: order of operations, what to gather before acting, when to stop and ask.
• Cross-check every skill against the environment’s tool table before pilot. A skill assuming tools outside the environment improvises or stalls.
• One canonical procedure per action. Overlapping skills produce inconsistent sessions and confused approvals.
• Verify each skill standalone in a test session: works, fails sanely on purpose (external slow, down, or empty), respects scope, gates where risky.
• Record each skill in the capabilities register: capability, layer, status (specified or verified), verification date, notes.

Per-skill verification checklist

• Works standalone in a test session.
• Failure modes observed once on purpose, output staying sane.
• Reaches only its intended systems and data.
• Approval behavior as designed: flows where smooth, gates where risky.
• Added to the capabilities register with kind, rationale, and verification date.

Composition rules, pitfalls, and the register format: Skills, tools and MCP. The registry surface is /ai/skills; treat the product UI there as the source of truth for fields beyond that guide.