From Software Repositories to Agent Skills: An Exploratory Empirical Study of Skillability in Open-Source Ecosystems

Abstract

AI agent ecosystems increasingly rely on reusable "skills," but deciding which open-source software projects are worth converting into skills remains largely ad hoc. We present an exploratory empirical study of 29,896 artifacts: 2,200 skills from the Clawhub marketplace and 27,696 GitHub repositories sampled from a larger filtered corpus. We operationalize skillability as a six-dimensional construct spanning task clarity, interface clarity, composability, automation value, deployment friction, and operational risk, and annotate artifacts with an LLM-based pipeline over metadata and README excerpts, validated on a 200-item human-coded subsample (87\% agreement within 0.5 points). The results reveal a large and structured conversion frontier. Marketplace skills score substantially higher than sampled GitHub repositories (3.75 vs.2.88; , Welch's -test , Cohen's ), and 35.8% of all analyzed artifacts satisfy our high-skillability threshold. High-skillability candidates concentrate in Data Retrieval & Search, Multimedia Content, and System Infrastructure, while raw skillability is effectively independent of repository popularity (Spearman ). Using skillability together with lightweight repository-quality signals, we identify 9,033 GitHub repositories as promising skill-conversion candidates. We position the paper as a scalable empirical foundation for repository-to-skill pipelines rather than as a finalized measurement paper. The contribution is a reusable rubric, a large-scale characterization of agent-facing software, and concrete evidence that open-source ecosystems contain enough high-potential repositories to make systematic and potentially batch-oriented skillification a realistic next step. The project website link is https://red0orange.github.io/repo2skill.

Keywords