216,938 skills, and only 114,565 come with a paper trail

Notes on SkillCenter: A Large-Scale Source-Grounded Skill Library for Autonomous AI Agents (arXiv:2607.07676) — Tianming Sha, Yue Zhao, Lichao Sun, Yushun Dong · July 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SkillCenter builds its library through a five-stage pipeline. Source acquisition feeds an LLM-based pre-filter called SkillGate, which screens raw material for actionability before any generation tokens get spent — auto-rejecting anything under 200 characters, then issuing a pass/maybe/fail verdict off a 0-10 suitability score. What survives moves through template-driven generation (four templates, keyed to source type) and an iterative improvement loop of up to three passes, each running a deterministic source grounding check that traces every retained claim back to an exact quotation in its source. A four-criterion publish gate — quality score, license compatibility, plagiarism ratio, placeholder density — catches roughly 3% of generated skills for falling under a 3-out-of-5 floor. Everything that clears it ships as an offline SQLite FTS5 database: keyword search, no embeddings, no vector store, no network call at query time.

The resulting library holds 216,938 skills across 24 domain bundles. 114,565 of them are pipeline-produced and source-grounded, drawn from 55,199 distinct documents — mostly PLOS and Nature journals plus ArXiv on the research side, GitHub, Stack Overflow, and web pages on the technical side. The remaining 102,373 are community skills, folded in from 90,984 harvested GitHub SkillMD files and 11,389 ClawHub marketplace listings. GPT-5.2 scores every pipeline skill 1 to 5; the average lands at 3.91, and 82.1% cluster at exactly 4 — a compression the authors attribute to known LLM-as-judge self-enhancement and verbosity biases, not genuinely uniform quality. Redundancy checks make the two halves' difference concrete: pipeline-generated domains sit under 0.01% duplicate content, while the harvested GitHub bundle carries a 26.9% near-duplicate rate from forks, templates, and copied boilerplate.