91.8% of public SKILL.md files are defective — and the defects are packaging problems you can lint for, not exotic attacks

arXiv 2608.08453·high signalXBlueskyLinkedInCopy link

A study of 138,133 public SKILL.md files from 20,556 repositories (arXiv 2608.08453, Aug 9) found 91.8% contain at least one detected defect (88.8–94.6% across thresholds), dominated by weak routing metadata, bloated or non-actionable bodies, and poor resource organization — because most skills were extracted from a single task or conversation rather than authored as reusable components. Skills written with spec-awareness had measurably fewer defects, while AI-marked skills showed more safety and portability problems. The actionable move: run a spec-aware generation workflow — specification-aware prompting, a lightweight linter, automated repair, then a safety gate — over your skill library instead of trusting one-shot extraction.