The bottleneck for capable agents was rarely raw reasoning — it was loading domain-specific procedure into context without bloating every prompt. Skills answer that with progressive disclosure: Claude sees only a one-line name and description until a task actually matches, then pulls in the full folder of instructions, scripts, and assets. This repository is Anthropic's reference collection plus the open specification for that format.
What Sets It Apart
- A skill is just a folder. The minimum is a
SKILL.mdwith YAML frontmatter (name, description) and Markdown instructions — no SDK, no runtime. That makes skills diffable, reviewable, and versionable like any other code. - Loaded by relevance, not always-on. Only the metadata stays resident; the body and bundled scripts load when the task warrants, so dozens of skills cost almost no baseline tokens.
- One format, three surfaces. The same folder works in Claude Code (as a plugin), on Claude.ai, and through the Skills API — authored once, run wherever Claude runs.
- Real document skills. The PDF, DOCX, PPTX, and XLSX skills here are the same capability behind Claude's file creation, not toy demos.
Great Fit / Look Elsewhere
Great fit if you want to encode a repeatable workflow — brand-compliant docs, an internal review checklist, a data pipeline — and have Claude apply it consistently without re-explaining each time; the template/ and spec/ folders give you a clean starting point. Look elsewhere if you need deterministic guarantees: skills are instructions Claude chooses to follow, not enforced code paths, and the document skills ship as demonstrations you should test before trusting on critical work.
