alchaincyf/darwin-skill

What it solves

Managing a large number of Agent Skills (defined in SKILL.md format) manually becomes impractical as the library grows. While traditional reviews focus on structural correctness (formatting and paths), they often ignore whether the skill actually produces the desired results. Darwin-skill provides a systematic way to optimize these skills by treating them as "trainable" assets that evolve through a measurable, iterative loop.

How it works

Inspired by Andrej Karpathy's autoresearch, the project implements a "ratchet mechanism" where skills are iteratively improved and only changes that result in a measurable score increase are kept; others are automatically rolled back via git.

The optimization process follows these steps:

1. Baseline Evaluation: The current skill is assessed against a 9-dimensional rubric.
2. Targeted Optimization: The system identifies the weakest dimension (weighted gap) and generates a specific improvement for that single dimension.
3. Independent Validation: Two independent sub-agents evaluate the new version to avoid self-evaluation bias. New judges are used each round to prevent anchoring effects.
4. Ratchet Application: If the score improves, the change is committed; otherwise, it is reverted.
5. Human-in-the-Loop: The process pauses at critical checkpoints (baseline, single-dimension optimization, and regression testing) for user confirmation.

Who it’s for

Developers and AI agent users who maintain a library of skills for tools like Claude Code, Codex, OpenClaw, Trae, or CodeBuddy and want to programmatically improve the quality and reliability of those skills.

Highlights

• 9-Dimensional Evaluation: Uses a rubric based on Microsoft Research's SkillLens, including failure mode encoding, actionable specificity, and high-risk action blacklists.
• Validation-Gated Edits: Aligns with the SkillOpt framework to ensure every edit is verified before being merged.
• Anti-Pattern Prevention: Explicitly forbids common pitfalls, such as having the same AI both edit and evaluate the skill.
• Git-Based Versioning: Uses git to ensure that the skill's quality only ever increases and never suffers from local degradation.