darwin-skill

About

darwin-skill, inspired by Andrej Karpathy's autoresearch, is a tool designed to optimize AI Agent Skills. It uses an autonomous experimentation loop to evaluate both the structural quality and actual performance of a Skill, preserving only changes that demonstrate measurable improvements. The system employs dual evaluation (static analysis and practical testing) and a 'ratchet mechanism' to ensure Skill scores only increase. It pauses at each optimization phase for user confirmation, enabling human-in-the-loop collaboration to enhance the continuous evolution of Skills.

Features

• Inspired by Andrej Karpathy's autoresearch for iterative Skill optimization
• Dual evaluation mechanism: combines structural scoring (60 points) and practical performance validation (40 points)
• Ratchet mechanism: only preserves improvements, automatically reverts regressions, ensuring scores only increase
• 8-dimensional evaluation system: comprehensively quantifies Skill quality
• Human-in-the-Loop: user confirmation required at each optimization phase

Supported Platforms