skill-audit

LLMs ignore documented rules during execution when SKILL.md files become bloated with duplicate instructions, inline code, or changelog noise. Auditing and pruning these files restores LLM rule effectiveness by eliminating patterns that overwhelm the model's context processing.

To audit a SKILL.md file, run a skill-audit script against it and its associated reference files. This identifies bloat patterns like duplicate rules, recency-bias accumulation, and missing Cardinal-N rule blocks, then generates a concrete cleanup plan to prune the file.

You should prune your SKILL.md file when a skill produces output that violates its own documented rules, or when the file exceeds approximately 500 lines without a deliberate reason. These are indicators of bloat that degrades LLM rule enforcement effectiveness.

Bloat patterns affecting LLM rule effectiveness in SKILL.md files include duplicate rules, inline CSS or code, historical changelog noise, recency-bias accumulation, and missing Cardinal-N rule blocks. Identifying and addressing these patterns streamlines rule enforcement.

Yes, after auditing a SKILL.md file and generating a cleanup plan, the skill-audit process can optionally execute the prune. This removes identified bloat patterns directly from the file to enhance LLM rule enforcement without requiring manual editing.