How is this list made?

Every number is from the GitHub API, read at 08:54 UTC on the day the page heading gives. Candidates are repositories tagged claude-code-skills, claude-skills, agent-skills, claude-code-skill, claude-skill and claude-code-plugin, and every repository linked from anthropics/claude-plugins-official, hesreallyhim/awesome-claude-code, travisvn/awesome-claude-skills, ComposioHQ/awesome-claude-skills and VoltAgent/awesome-agent-skills; the lists find candidates and never decide a rank. The 15 with the most stars that pass the checks below are the list, and each tab orders the same 15.

How each tab ranks

All time is total stars. This week and this month score the last 7 and 30 days: forks gained 50%, commits 20%, active contributors 20% and a release 10%, each as a share of the list's highest value in that window. The list is read once a day, and once a daily snapshot is old enough to measure the window, stars gained join: stars gained 30%, forks gained 30%, commits 15%, active contributors 15% and a release 10%. A rank's movement is against the previous day's ranking; until there is one, every entry is new.

GitHub's API gives no date for a star, so the curves are forks, which carry the day each was made. They read the oldest-first fork list at 12 evenly spaced positions, and a window's forks are counted exactly. Commits are to the default branch, and active contributors are the people who made them, bots left out. A fork since deleted is in none of these.

What is left out

Forks, archived repositories, anything with no commit in six months, anything that ships no SKILL.md file, star farms (at least 1,000 stars, with forks under 1% and watchers under 0.1% of them), and repositories with no licence GitHub can detect. One exception to the last: a repository with no licence at its root stays when at least 90% of its skill folders carry a licence file of their own, and its licence then reads "Per skill". A repository is also left out by hand when its README shows it is something else, or says too little to describe it.

What each entry does, its type and its category are written from its own README and the files it ships.

Left out, with more stars than the last one listed

• morluto/rea, 221,844 stars: an MCP server for reverse engineering; its skill is the workflow for driving it
• multica-ai/andrej-karpathy-skills, 219,602 stars: no licence GitHub can detect, at its root or in its skill folders
• anthropics/claude-code, 150,722 stars: no licence GitHub can detect, at its root or in its skill folders
• paperclipai/paperclip, 101,788 stars: an app, a server and dashboard for running teams of agents
• nexu-io/open-design, 101,045 stars: a design app, of which skills are one part
• shareAI-lab/learn-claude-code, 78,546 stars: a course on building agent harnesses
• ComposioHQ/awesome-claude-skills, 77,055 stars: no licence GitHub can detect, at its root or in its skill folders