Skills
Generate agent skills from source material instead of authoring them: +19.9 points over running with no skill
SkillAlchemy tackles skill creation when there are no human authors, model priors, or execution traces to copy from, taking an underspecified brief plus a source-access spec, surfacing implicit requirements through contrastive evidence, admitting candidate procedures only where the evidence supports their scope, then compiling the admitted content into a grammar-guided skill package. Across 87 SkillsBench v1.1 tasks it beats no-skill execution by 19.9 points and the strongest automated baseline by 8.6, reaching parity with human-curated skills. The admission step, deciding how broadly each derived procedure is justified, is the transferable idea for anyone writing skill files by hand.
↳ Follow the thread