Agent Skills
Also called Skills, Agent skill, SKILL.md
Folders of instructions, scripts and resources that an AI agent loads only when a task needs them, to do that kind of task well.
What are agent skills?
Agent skills are packages of know-how for AI agents. Each skill is a folder containing a main instruction file (usually called SKILL.md) and, optionally, scripts, templates and reference documents. A skill might teach an agent to fill in your company's report template, follow your brand guidelines or work with spreadsheets.
Skills started at Anthropic and are now published as an open format that other agent tools can support.
How does it work?
The agent starts with only a short name and description of each installed skill, which takes up very little of its context window. When a task matches a skill, the agent reads the full instructions, and opens the extra files or runs the scripts only if it needs them. This "load on demand" approach is a form of context engineering: an agent can have many skills installed without its context filling up.
What is it commonly confused with?
- MCP servers. An MCP server connects an agent to an outside tool or data source. A skill teaches the agent how to do a kind of task. They work well together: a skill can explain how to use the tools an MCP server provides.
- A system prompt. A system prompt is always loaded; a skill is loaded only when it's relevant.
Why does it matter?
Skills let teams write down a process once and have agents follow it consistently, without pasting long instructions into every conversation. Because a skill can include scripts the agent runs, treat skills from other people like any software you install: read them first and use ones from sources you trust. Read MCP vs agent skills for when to use each.

