Blog
Studies from the registry's own crawl of public GitHub, each built on a count nobody else publishes, with the method and the data beside it. For how-to answers, see the guides; for the full data, the reports. New posts are in the feed.
. We followed 31,009 SKILL.md, AGENTS.md and other agent files on GitHub for four weeks: how many changed, and how often a skill's description moved.
. We read 2,536 CLAUDE.md files on GitHub: how many only point to AGENTS.md, how many run past the 200-line target Claude Code sets, and what the rest hold.
. We read 4,687 AGENTS.md files on GitHub: which recommended sections they have, and how many tell the agent how to check its work.
. We read 70,893 agent skill folders on GitHub: how many are SKILL.md alone, how many ship code an agent can run, and what else sits beside it.
. We read 73,322 SKILL.md files on GitHub: how many descriptions say when to use the skill, and what else the files add.
. We ran the DESIGN.md format's own linter on 1,162 DESIGN.md files from GitHub: how many carry design tokens at all, and what it flags in those that do.
. We read 438 llms.txt files on GitHub against the llms.txt proposal: how many open with an H1 and a summary, and how many keep their sections as link lists.
. We read 557 Cursor rule files on GitHub: how many are .mdc or legacy .cursorrules, which rule types they use, and which patterns Cursor ignores.
. We read 1,203 folders on GitHub that hold both AGENTS.md and CLAUDE.md: how many import, symlink or copy one file, and how many keep two that differ.