skill-optimize

# Skill Optimize ## Purpose / When to Activate Activate when: - A skill needs a baseline quality measurement (no ledger.yaml exists yet) - A SKILL.md has been modified and a before/after regression check is needed - `friction-retrospective` flags a skill as a recurring friction source - An eval cycle is needed after a manual skill update This skill formalizes the Skill Optimize protocol referenced by `eval-run` (SKILL-CRS-025). It runs the full evaluate-analyze-improve loop with mandatory human gates at every modification step. **Scope:** Any GAAI skill with measurable quality criteria — not limited to content-production skills. The inline Skill Optimize protocol in `discovery.agent.md` remains the agent's orchestration logic; this skill provides the structured execution procedure. --- ## Process ### Step 1 — Eval authoring If no `evals.yaml` exists for the target skill: 1. Read the target `SKILL.md` in full. 2. Identify measurable quality criteria from the `Quality Checks` section. 3. Author `evals.yaml` following the `evals-format.md` spec (see `eval-run/references/evals-format.md`). 4. Include a minimum of 5 assertions with a mix of `code` and `llm-judge` types. 5. Store the file at `{skill-dir}/eval-corpus/evals.yaml`. **HUMAN CHECKPOINT:** Present the drafted `evals.yaml` for validation. Do not proceed until approved. If rejected, revise based on feedback and re-present. ### Step 2 — Corpus generation If no corpus outputs exist in `{skill-dir}/eval-corpus/`: ...