Coding agents can be guided by instruction files stored in a repository. We conducted a census of popular, active GitHub repositories across thirteen languages, treating each repository’s root AGENTS.md and CLAUDE.md as a single instruction set. The survey found that AGENTS.md and CLAUDE.md appear in a substantial share of projects and in many of the most‑starred ones; when both files exist, CLAUDE.md usually points to AGENTS.md. We applied model‑assisted annotation together with human review to analyse the statements. The annotations reveal a practical blend of project knowledge and rules about code, build, and workflow. Collaboration rules enumerate concrete requirements such as test plans and commit formats, yet roughly half of the directives provide no rationale. Policies on AI attribution diverge: commit‑message rules tend to prohibit attribution, whereas pull‑request rules more often require disclosure. About one‑third of the instruction sets carry a generator‑associated signature, a tool‑inserted block, shared text, or an AI‑marked creation commit; four‑fifths are edited after creation. Prespecified tests link repository popularity to pull‑request and AI‑contribution rules. Exploratory analyses show that activity correlates with process‑oriented content and that LLM‑ and agent‑related projects are over‑represented among adopters. The models agree with each other far more closely than with the human reviewer’s topic assignments, highlighting a caution for automated annotation. We release the annotations, taxonomies, and pointers to the analysed file versions.
Review: This work delivers the first systematic dataset of open‑source instruction files, offering empirical insight into how AI agents are governed in software development.