skill-improver

Automating iterative skill improvement uses a greedy hill-climbing autoresearch loop that scores, hypothesizes, mutates, and re-evaluates SKILL.md files to raise overall quality across multiple dimensions.

Blind validation in an autoresearch loop reduces scoring bias by re-evaluating mutated skill documentation without prior context, ensuring objective quality checks before persisting changes to the backlog.

Scoring Claude Code skills against a rubric involves evaluating SKILL.md frontmatter and content across 10 dimensions, generating improvement hypotheses, and applying atomic changes per iteration.

Batch workflows support improving multiple skill directories at once by using batch-mode orchestration to apply the autoresearch loop across many skills sequentially within the .claude/skills structure.

Yes, you need local skill directories under .claude/skills along with the references and scripts structure, as the autoresearch loop requires direct file access to perform scoring, mutations, and backlog updates.

Available modes for optimizing Claude Code skill documentation include improve, score, freshen, trigger, philosophy, and batch, each targeting specific aspects of the autoresearch and quality enhancement workflow.