SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for Real-World LLM Agents

Short summary

SkillCorpus aggregates ~821,000 crawled SKILL.md files into 96,401 curated skills organized by a 16-class taxonomy with quality facets for utility, robustness, and safety. Paired with a fine-tuned retrieval-and-selection stack, it yields consistent gains across three benchmarks, with the largest improvement on SkillsBench (+7.5 pp). The work provides the first end-to-end account of when a curated community skill corpus improves real agent tasks and where it does not.

• •SkillCorpus filters 821K crawled SKILL.md files into 96K curated skills with 16-class taxonomy and quality facets
• •Retrieval-and-selection stack yields consistent gains across SkillsBench (+7.5 pp), GDPVal, and QwenClawBench
• •First end-to-end evaluation of open-source skill ecosystem benefit on real-world agent tasks

Generated with AI, which can make mistakes.