| name | skill-optimizer |
| description | Improve a skill through benchmarked, validation-gated edits instead of ad hoc prompt polishing. |
Skill Optimizer
Use this skill when the user asks to optimize, tune, or improve an existing skill against examples or a benchmark.
This port is adapted from garrytan/gbrain at commit 814258dda67945ffec9457a1e73980e947b7e462.
Workflow
- Pick the target skill and read its
SKILL.md.
- Establish an evaluation set before editing:
- existing examples, tests, or benchmark JSONL
- user-provided success criteria
- at least a small held-out set for bundled skills
- Make the smallest body-only change likely to improve behavior.
- Preserve frontmatter routing fields unless the user explicitly asked to change routing.
- Validate before accepting:
- repo tests
- relevant skill-specific checks
- manual comparison against held-out tasks
- Record what improved, what was rejected, and remaining risk.
Codex Adaptation
- Prefer deterministic local tests and review tasks over upstream
gbrain skillopt unless that CLI is installed and the user approved its cost/runtime.
- For bundled skills in this repo, default to proposed edits plus validation rather than automatic mutation loops.
- Do not call external model optimization services unless the user explicitly approves the cost and network use.
Guardrails
- Do not optimize against a vague preference; write the benchmark first.
- Do not mutate frontmatter in a quality-tuning pass.
- Do not accept a change that only improves the training examples while regressing held-out tasks.
- Do not present an unvalidated rewrite as an optimization.
