A simple format for a simple concept: that’s how Anthropic presents Claude Skills.
It isn’t so much a feature – the American group even avoids that word – as a particular way of introducing context. Specifically, via Markdown files and any associated resources (code, templates, documentation, etc.).
The file in question (SKILL.md) contains a YAML header naming the skill and describing it. This approach paves the way for what Anthropic calls a “progressive disclosure,” so Claude does not overload its context window.
The model does not actually access the skills right away. It first loads their name and description into its prompt framework, then enables them or not depending on the tasks it has to accomplish.
Building on AGENTS.md
Claude Skills follows in the footsteps of AGENTS.md, a “readme for coding agents” that emerged under the impetus of Google, Cursor and OpenAI, among others. It adds, however, a tree-like structure, with SKILL.md able to call other Markdown files located in the same folder.
While the mechanism appears reproducible with other providers, its current implementation is tied to the Anthropic ecosystem. It relies on the Bash tool to read Markdown files and to optionally execute any associated scripts.
Any skill that is activated enters Claude’s context window (order of magnitude: up to roughly 5,000 tokens, according to Anthropic, with the name and description consuming about 100 tokens).
Finding Complementarity with MCP
The system has been in use for several weeks on Claude.ai, enabling document creation (Word, Excel, PowerPoint, PDF). It is available on the Pro, Max, Team and Enterprise plans. A designer is on hand to create skills… using the same Claude. They can then be imported in .zip format via the settings. They are user-specific.
Using Claude Skills on the API Messages requires three headers: skills-2025-10-02 (feature active), code-execution-2025-08-25 (allowing skills to run in the code executor) and files-api-2025-04-04 (enables file uploads and downloads).
Skills are uploaded via the /v1/skills endpoint. They are accessible to the entire organization. To invoke them, they are included in the container parameter by specifying their identifier, type and, optionally, their version. Up to eight can be included per request.
The skills are also available with Claude Code, including in the form of a plug-in. They can be personal or shared.
Anthropic says it is considering how to complement MCP, to “teach agents more complex workflows involving external tools.” It also envisions a day when these agents could autonomously create their own skills.
