skill-evolve

You systematically improve Claude Code Skills by running observation-driven evaluation cycles on real prompts, recording outcomes, extracting repeating patterns, and applying small, verifiable changes to raise quality.

Observation-driven skill evolution is an iterative improvement process that observes real prompt outcomes, extracts repeating patterns, prioritizes high-impact changes, and validates them through documented evolution-log rounds for reproducibility.

To start an observation cycle, provide a target skill path, read its SKILL.md file containing the YAML frontmatter, and initiate the first guided observation cycle documenting prompts and recording outcomes.

Yes, you need a Skill Unit defined by a SKILL.md with YAML frontmatter containing name and description, alongside optional scripts, references, and assets directories for operational assets required by the evolution workflow.

Yes, the evolutionary workflow manages rounds, maintains an evolution-log, and applies versioned changes to ensure continuous improvement and reproducibility of pattern-based modifications across observation cycles.

Pattern-based improvement relies on running real prompts to extract repeating patterns, meaning insufficient prompt variation or lack of documented validation steps can limit the extraction of high-impact, verifiable changes.