Measured, not claimed
The scan, lift, reviewer and library numbers come from runs you can repeat from the repository. The token week on the landing page is one developer's own logs. Small samples are labelled as small.
- We scanned 511 skills from the skills.sh leaderboard repos: what we found
Scan of 511 unique skills from 34 GitHub repos (2026-10-29): 96 ship scripts, 85 make network calls, 0 confirmed malicious. Method and limits included.
- Do agent skills help? Lift evals with and without the skill
Same task, same model (Claude Haiku), with and without the skill installed: four skills, 2 runs per condition, fixed checks and no AI judge.
- How accurate is the Skillhawk reviewer? A false-positive corpus
We graded six well-made skills (ibelick/ui-skills, MIT) before and after fixing the reviewer's false positives: grades moved from D, C, F to A and B.
- Library grades: 277 public skills reviewed and security-checked
Grade distribution and security status for 277 public agent skills, judged 2026-11-17 with skillhawk review and skillhawk secure.