What problem does it solve? Agent Skills can drift over time: descriptions route incorrectly, SKILL.md files bloat, and edits overfit to a single failure. This Skill provides a measurable loop—baseline, evaluate, bounded edit, re-validate—so skill quality improves without destroying working rules. ## Core Features & Use Cases - Layered evaluation stack: Runs validator gates, rule-based YAML eval cases, Codex plugin-eval static analysis, human prompt suites, and SkillOpt-style bounded edit loops, choosing the cheapest layer that answers the question. - T1/T2 quality gates: Applies T1 behavior-critical CI checks (pnpm run eval with evals/cases/*.yaml) versus T2 structural validate-only checks per ADR 0011 and ADR 0027. - Held-out gating: Splits prompt suites into train and held-out sets, keeping an edit only if held-out results improve, and logs outcomes in references/evals.md. - Use Case: After reviewers report a skill's description is vague and routing fails, run layer 0 validation, add rule-based eval cases, apply a bounded edit under 10% line churn, and re-run held-out prompts before merging. ## Quick Start Ask the agent to improve a specific skill after routing eval failures and suggest one bounded quality fix with a held-out gate.
