Anthropic Agent Skills
★ 157.5k · GitHub →Anthropic's official Agent Skills — folders of instructions and scripts Claude loads on demand.
What is Anthropic Agent Skills?
anthropics/skills is the official repository for Agent Skills — Anthropic's mechanism for packaging procedural knowledge that Claude loads only when a task calls for it. A skill is just a folder with a SKILL.md file: YAML frontmatter declaring a name and a description of when to use it, followed by markdown instructions, optionally accompanied by scripts and reference files. The agent reads the description index up front and pulls in the full skill body on demand, so expertise scales without permanently bloating the context window.
The repo carries two tiers of content. The document skills — docx, pdf, pptx, xlsx — are the production-grade ones: they are what powers Claude's own file creation, pairing instructions with Python scripts, and are source-available rather than open source. Around them sits a set of Apache-2.0 example skills covering creative, development, and enterprise workflows, explicitly framed as demonstrations to learn the format from.
The important conceptual point is that skills are complementary to MCP, not competing with it. MCP gives an agent new capabilities — APIs it can call; a skill gives it knowledge and procedure — how to do something well with the capabilities it already has, often just Bash and file editing. Skills install via the Claude Code plugin marketplace, upload to claude.ai on paid plans, or attach through the Claude API, and the SKILL.md format itself is a published open spec.
Key capabilities
- ▸ SKILL.md format: a folder with YAML frontmatter + markdown instructions, optionally bundling scripts and references
- ▸ Progressive disclosure — only skill descriptions sit in context until a task triggers a full load
- ▸ Production document skills (docx, pdf, pptx, xlsx) — the same ones behind Claude's own file creation
- ▸ One-command install in Claude Code via the plugin marketplace; also usable on claude.ai and the Claude API
- ▸ Open specification, so the format works as a template for team-specific skills beyond this repo
When to use it
Install the document skills if you generate Word, PDF, PowerPoint, or Excel files with Claude — they are the strongest concrete win here. Study the example skills when you want to encode your team's procedures (deploy runbooks, review checklists, house style) into something the agent reliably follows. Do not reach for skills when the agent needs a new external capability — that is MCP's job — and note the whole system is Claude-ecosystem-specific: if you run agents on other model stacks, the SKILL.md files are readable prose but nothing will auto-load them.
Alternatives
Frequently asked questions
How do I install Anthropic skills in Claude Code?
Add the marketplace with `/plugin marketplace add anthropics/skills`, then install a bundle such as `/plugin install document-skills@anthropic-agent-skills` (example-skills works the same way). After installation, just mention the task — "turn this data into an xlsx" — and Claude Code loads the matching skill automatically based on its description.
What is the difference between Agent Skills and MCP servers?
MCP servers add capabilities: running processes that expose new tools the agent can call, like a browser or the GitHub API. Skills add knowledge: markdown instructions and scripts teaching the agent how to perform a task well with capabilities it already has. They compose naturally — a skill can describe the procedure for using an MCP server's tools effectively.
Can I write my own skill using the SKILL.md format?
Yes — that is the intended use. Create a folder containing SKILL.md with YAML frontmatter (a lowercase hyphenated name plus a description of what it does and when to use it), followed by markdown instructions; add helper scripts alongside if needed. The description matters most, since it is what the agent reads when deciding whether to load your skill. The format is an open, published spec.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →