name: skill-optimizer
description: Improve another skill by running repeated test prompts, scoring outputs against binary eval criteria, and proposing tighter SKILL.md instructions. Use when the user wants to auto-improve a skill, build an eval suite for a skill, adapt Karpathy-style autoresearch to prompt/skill optimization, or run iterative optimization on a skill such as workout-log.

Skill Optimizer

Use this skill to optimize another skill's instructions with an eval-driven loop.

Read references/autoresearch-pattern.md for the minimal pattern lifted from Karpathy's autoresearch repo.
Read references/eval-design.md when designing binary evals.
Read references/workout-log-evals.md when the target skill is workout-log.
Use scripts/start_workout_log_eval.sh to initialize a repeatable workout-log eval run and store artifacts under assets/workout-log-eval-run/.
Use scripts/score_template.md as the scoring sheet for each pass.
Use scripts/append_result.py to append summary scores to the TSV after a pass.

Core loop

1. Pick one target skill only.
2. Read the target SKILL.md and any required references.
3. Build a small test set of realistic user prompts.
4. Define binary eval questions for each prompt.
5. Run the target skill against the tests.
6. Score the outputs.
7. Tighten the target skill instructions.
8. Re-run the same tests.
9. Keep the improved version only if the score is better or meaningfully simpler at equal score.

Rules

• Keep evals binary whenever possible.
• Prefer 4-8 eval checks per test prompt.
• Test with realistic prompts, not idealized prompts.
• Do not optimize multiple skills in one pass.
• Preserve the target skill's purpose; improve reliability, not scope creep.
• If a result is ambiguous, mark the eval as failed instead of inventing a pass.
• Favor simpler prompt edits over bloated prompt edits.

Output shape

When reporting an optimization pass, include:
- target skill
- test prompts used
- eval criteria
- baseline score
- revised score
- what changed
- whether to keep the revision

Recommended first target

Start with workout-log before trying browser-heavy skills.
It is easier to evaluate, less noisy, and failures are more obvious.