darwin-skill, Agent Skill Optimization

darwin-skill is an optimization system that evaluates SKILL.md used across multiple coding agents, modifies only one dimension at a time, and verifies changes through tests. Inspired by Andrej Karpathy's autoresearch, it applies a ratchet structure that only commits improved changes.

The core loop is as follows.

• Evaluate the current skill
• AI generates one improvement proposal
• Regression test with test-prompts.json
• Two independent sub-agents re-evaluate
• If the score rises, git commit; if it falls, git revert
• Human checks the diff and score every round

darwin-skill 2.0 reflects Microsoft Research's SkillLens and SkillOpt research, expanding the evaluation criteria to 9 dimensions, out of 100 points. Beyond structural completeness, it evaluates execution performance, failure mechanism encoding, actionable specificity, and a destructive command blacklist, with the highest weight placed on execution performance.

To reduce self-evaluation bias, the modifying agent does not score its own work directly, and a new judge is used every round. The post explains, citing SkillLens figures, that LLM self-evaluation accuracy is only 46.4%.

If a round's improvement is less than 1 point, the process terminates early to prevent unnecessary padding. In a public example, the huashu-gpt-image skill score rose from 80.8 to 91.65.