Test of 20 top agent skills: the most-installed one made answers worse

Driftproof ran the top 20 skills from the 17 October skills.sh snapshot on identical tasks with and without each skill, using Claude Sonnet 5.5 and Opus 5.5 with Haiku 5.5 as judge. 11 skills clearly helped on both models, while the most-installed one, find-skills, hurt: 0.75 versus 0.92–0.95 without it.

• 11 of 20 skills clearly helped on both models
• find-skills (vercel-labs) hurt: 0.75 vs 0.92–0.95 without it
• hyperframes-registry lifted Sonnet scores from 0.43 to 0.98
• 10 of 40 results were inconclusive due to variance