• Sources: primary
  • Summary: Every skill change must pass structural and lint checks, a required semantic version bump, a validated eval suite with a fresh report tied to the latest commit, link checks and an AI review agent, then human approval. The AI reviewer pushes rigid CLI sequences into scripts, caps SKILL.md at 500 lines and blocks chained references, which evals showed agents often fail to follow. Cheap deterministic checks run first so failing changes never trigger the costly LLM gates.
  • Why it matters: It is a concrete template for treating agent instructions as versioned, evaluated code with atomic eval expectations.