From Registry to Repository: How AI Agent Skills Are Written, Adapted, and Maintained

Jul 16, 2026

Upload by Ivy

AI coding agents increasingly rely on “skills” as structured context bundles, typically stored as SKILL.md files with YAML headers and Markdown bodies. Central registries like skills.sh now host tens of thousands of skills, yet skills are often treated as agent capabilities rather than evolving software artefacts. This study analyzes 18,463 registry skills and 23,199 personal-use skills from 5,876 repositories, mapping 3,709 reuse links, classifying them via SWEBOK areas, and coding modifications to reveal how reuse and maintenance shape agent behavior.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Introduction

Skills as structured context bundles

Motivation: limitations of prior research

Research questions and empirical approach

From Registry to Repository: How AI Agent Skills Are Written, Adapted, and Maintained Haoyu Gao The University of Melbourne Victoria, Australia [haoyug1@student.unimelb.edu.au](haoyug1@student.unimelb.edu.au) Jai Lal Lulla Singapore Management University Singapore [jailal.l.2025@phdcs.smu.edu.sg](jailal.l.2025@phdcs.smu.edu.sg) Hong Yi Lin The University of Melbourne Victoria, Australia [tom.lin@student.unimelb.edu.au](tom.lin@student.unimelb.edu.au) Sebastian Baltes Heidelberg University Heidelberg, Germany sebastian.baltes@uni-heidelberg.de Christoph Treude Singapore Management University Singapore [ctreude@smu.edu.sg](ctreude@smu.edu.sg) Mansooreh Zahedi The University of Melbourne Victoria, Australia [mansooreh.zahedi@unimelb.edu.au](mansooreh.zahedi@unimelb.edu.au) arXiv :2607 .009 1 1v2 [ cs . SE] 6 Jul 2026 Abstract—AI coding agents increasingly rely on skills: structured context bundles, typically a SKILL .md file with a YAML header and Markdown body, loaded on demand for domain knowledge, workflows, and scripts. Public registries such as [skills.sh](skills.sh) now host tens of thousands of skills, making them an emerging unit of reuse in agent-based software engineering. Yet skills have largely been viewed as agent capabilities rather than software artefacts whose content and evolution shape agent behaviour. We present the first empirical study of AI agent skills as engineered artefacts that are authored, reused, customised and maintained, across public registries and personal-use repositories. We mined 18,463 skills [from](from skills.sh)[ skills.sh](from skills.sh) and 23,199 personaluse skills from 5,876 GitHub repositories, identifying 3,709 reuse links. LLM-based classification into SWEBOK knowledge areas (KAs) shows Software Construction dominates alongside a long tail of specialised areas. A thematic analysis of 180 skills identifies six content categories. Qualitative coding of 444 modifications reveals six themes, of which reworking operational specificationsand adapting knowledge and resources are the primary target of change. Our findings show that reuse is largely a one-time copy operation: most reused skills remain near-verbatim, 53% are never modified after adoption, and subsequent local maintenance is overwhelmingly additive. Customisation primarily adapts skills to local environments, whereas evolution accretes new inline domain knowledge. Across both, a stable behavioural contract—how a skill interacts with users, monitors runtime state, andrecovers from failures—remains almost untouched. These results suggest maintenance effort should focus on project-specific bindings, and that registries and tool support should enable consolidating the domain knowledge skills re-author in isolation. I. INTRODUCTION Recent advances in large language models (LLMs) have given rise to coding agents that participate directly in software development, contributing to code generation [1], code review [2], program repair [3], test generation [4], and documentation [5] . To extend these agents into specialised tasks while keeping the context window manageable, the community has converged on the concept of skills: structured context bundles loaded on demand, supplying just-in-time domain knowledge, prescribed workflows, and optional companion scripts or reference documents. A skill typically consists [of a](of a SKILL.md)[ SKILL.md](of a SKILL.md)[ ](of a SKILL.md)file with a short YAML header and a Markdown body, packaged in a folder alongside optional assets. Skills reach an agent through two distinct channels. Some are published to centralised registries such as [skills.sh](skills.sh) [6], which already hosts tens of thousands of skills, whilst others are authored by developers for the individual repositories and maintained alongside the source code as project-specific infrastructure. This mirrors the distinction between broadly reused third-party libraries and locally authored, project-sp

FAQs

What are AI agent skills in this document?

Skills are structured context bundles, usually packaged as SKILL.md files with a YAML header and a Markdown body, optionally accompanied by scripts or assets. They are loaded on demand to provide domain knowledge, workflows, and references.

How did the study collect data about skills?

The authors mined 18,463 skills from the public skills.sh registry and 23,199 personal-use skills from 5,876 GitHub repositories. They then recovered 3,709 reuse linkages connecting published skills to locally adapted versions.

What do the results say about reuse and later maintenance?

Reuse is largely a one-time copy operation: most reused skills stay near-verbatim, and 53% are never modified after adoption. When changes occur, maintenance is mainly additive, focusing on adapting bindings and incorporating new inline domain knowledge while leaving a stable behavioural contract mostly untouched.

Jul 16, 2026

Upload by Ivy