Skip to content
EnglishFrançais

Agent skills

Agent skills are packages of instructions that give an AI agent a specific capability: a folder containing a SKILL.md file, with a name, a description and step-by-step instructions, plus optional scripts and reference files. The agent loads a skill only when a task matches its description. Anthropic created the format and released it as an open standard in December 2025.

Definition and how it works

Skills are a building block of agentic AI: they write down how a task is done, the way a team documents its own methods, so that an agent can apply that method without being told again in every conversation. According to the open specification published at agentskills.io, the SKILL.md file carries at least a name and a description in its header, followed by the instructions, and the folder can also hold scripts, reference documents and templates.

Skills rely on progressive disclosure, in three stages. At startup, the agent reads only the name and description of each available skill. When a request matches a description, it loads the full instructions, and it opens bundled files or runs scripts only when those instructions call for them. Anthropic's documentation puts the cost of the first stage at about 100 tokens per skill, which is why an agent can keep many skills available without filling its context window.

Anthropic announced agent skills for Claude on 16 October 2025 and published the format as an open standard on 18 December 2025. The specification states that a growing number of agent products have adopted it, so a skill written once can be reused across compatible agents.

Agent skills, MCP and prompts: the differences

The three are often confused because each one changes what an agent does, but they act at different levels. A prompt is an instruction given in one conversation and lost at the next, whereas a skill is a reusable procedure that the agent loads whenever a task calls for it. The Model Context Protocol (MCP), for its part, is an open protocol that connects an AI application to external systems, such as a database, a file store or a CMS, so that it can read their data and call their functions.

Prompt, agent skill and MCP server
  Prompt Agent skill MCP server
What it provides An instruction for the current request A method: steps, rules, examples, scripts Access to an external system: data, actions and prompt templates
How long it lasts One conversation Every conversation where it applies As long as the connection is configured
Format Free text A folder with a SKILL.md file A server that follows the MCP specification

Skills and MCP servers work together, since a skill can tell the agent which MCP tools to call and in what order while the MCP server provides the access. An agent skill is also distinct from "AI skills" in the sense of the competencies people need to work with AI, such as writing prompts or checking outputs, which is the other meaning search engines return for the term.

Agent skills in practice

For a marketing or digital team, a skill is a way to write a method down once: an editorial style guide, a review checklist, a publication workflow with its validation steps. The French open data platform data.gouv.fr, for example, publishes a skill that documents how to use its APIs alongside a separate MCP server that gives access to them, a clear illustration of the division of roles, although the platform still describes its skill as experimental.

Because a skill can contain code that the agent runs, installing one carries the same risk as installing software. Anthropic advises using skills only from trusted sources and auditing every file of a skill from an unknown source, since a malicious skill could direct the agent to call tools or run code that does not match its stated purpose. In an organization that connects agents to its content management system, that review belongs in the same governance as any other integration.

Frequently asked questions

Resources that cover this term