SkillCorpus consolidates 821K open skills, boosting LLM agents by 7.5%

The open-source ecosystem for LLM agent skills is booming but fragmented—thousands of reusable procedural knowledge packages exist, yet they are redundant, uneven in quality, and lack standardized evaluation. SkillCorpus tackles this head-on. The team crawled ~821,000 skills from public repositories, ran them through a multi-stage pipeline (deduplication, quality filtering, safety checks), and ended up with 96,401 high-quality skills organized into a 16-class taxonomy. Each skill is also scored on three key facets: utility, robustness, and safety. To make these skills actually useful, they built a fine-tuned retrieval-and-selection engine that picks the right skills for a given task.

In end-to-end tests across three benchmarks—SkillsBench, GDPVal, and QwenClawBench—using two different agent harnesses and two open-source backbone models (plus a frontier robustness check), SkillCorpus consistently improved performance. The biggest lift was on SkillsBench, where the curated corpus delivered a +7.5 percentage point gain over baseline. The authors also identified two key boundaries that limit further gains: a coverage boundary (the corpus doesn't yet cover all possible skills) and a harness boundary (the agent framework itself may not fully utilize all retrieved skills).

SkillCorpus is, by the authors' account, the first end-to-end demonstration that a curated, retrieval-served community corpus can improve real-world agent tasks—and where it still falls short. The dataset, models, and code are promised upon acceptance. For the growing field of LLM agents, this provides a systematic foundation for building, sharing, and trusting open skills.