Audit your SKILL.md files: ~60% of skill body is non-actionable filler you're paying for every invocation

arXiv·low signal

SkillReducer's analysis found over 60% of typical skill-file content is prose that doesn't change agent behavior, and reference files can inject tens of thousands of tokens per call. Mechanically compressing skills yielded a mean 77.5% token reduction (95.8% on skills over 10k tokens) without losing capability — a strong argument to strip narrative from your skills and push detail into on-demand referenced files.