/ agent-skills
GitHub →

Agent skills, curated and pinned

Skills for Claude Code, OpenCode, Codex and the other agents that read SKILL.md. Own skills and reviewed upstream ones, each pinned to a commit, with its context cost in plain numbers.

How a skill spends context

Agents load skills in three steps. Only the first costs something in every session, which is why descriptions stay short and bodies stay lean.

SKILL.md body, loaded when the skill fires over the 5,000-token guideline guideline

Skills

Own skills are written here. Vendored skills are byte-for-byte copies of upstream at a pinned commit; a weekly job proposes updates as pull requests with a review report.

From publishers

These carry more than skills (hooks, a language server, sub-agents, commands), so they install as plugins straight from their publishers. Their pure skills are vendored above.

PluginWhat it is forWhy a plugin

Reference corpora

dev-references searches these repositories through a self-hosted index. Only this list is published; the content is cloned onto the machine that builds the index.

CorpusUse it forLicense

Activity

Recent commits: new skills, reviewed upstream bumps, catalog changes.