Agent Skills are expertise packs that Anthropic designed for Claude: a folder holding instructions and files for one kind of work, which Claude opens only when it needs them. In Chinese, they are often called Claude Skills.
What are Agent Skills
Picture hiring a new coworker. You wouldn’t read them every rule on day one. You’d hand them an onboarding handbook and let them flip to the right page when they need it.
A Skill is that handbook. The folder holds instructions, small programs, and reference material. With one installed, a general-purpose Claude can act like a specialist in a particular area, such as formatting reports the way your company does.
Every Skill needs an instruction file called SKILL.md, and the top of that file must fill in two fields: a name and a description. The description should say clearly what the Skill does and when to use it.
Anthropic has already built a few common ones, called prebuilt Skills, covering PowerPoint, Excel, Word, and PDF. You can also make your own, called custom Skills. Both work the same way: once a Skill is in your environment, Claude uses it automatically when it fits.
How Claude Skills work: opened only when needed
The key idea is loading in stages, which the official docs call progressive disclosure. Claude doesn’t read everything into its head at the start. It loads things in order, only as needed.
- It remembers only the name and description. Until a Skill is used, Claude sees just these two things, so installing many Skills adds little weight.
- It reads the instructions once a task matches. Only when your request fits a Skill’s description does Claude read the body of
SKILL.md. - It fetches other files only if needed. Templates and details that aren’t needed stay in the folder and don’t use up the amount Claude can read at one time. A small program just runs and hands Claude only the result.
How Claude Skills differ from a regular prompt
A prompt is an instruction you give in a conversation, and it usually covers just that one task. Skills load on demand, so you don’t have to repeat the same approach in every conversation.
Here’s an analogy. Telling the delivery app “leave it at the door, no utensils” on every single order is a prompt. Saving that as a default note that gets attached automatically every time is like a Skill.
The official docs list three benefits: Claude becomes specialized for specific tasks, you repeat yourself less, and you can combine several Skills for multi-step work.
Where you can use Agent Skills
| Where | Prebuilt Skills | Custom Skills |
|---|---|---|
| claude.ai | Turn on automatically when you create documents, no setup | Upload as a zip file; Pro, Max, Ultra, Team, and Enterprise plans only, and code execution must be on |
| Claude API (the interface for developers) | Available, used together with the code execution tool | Once uploaded, everyone in the same workspace can use them |
| Claude Code (the coding version) | Document-type prebuilt Skills are not available | Put them in the designated Skills folder on your computer, no upload needed |
One easy mistake to make: custom Skills don’t sync across platforms. A Skill you upload to claude.ai isn’t visible to the API, and the reverse is also true, so you manage each place separately.
What to check before using Agent Skills
A Skill contains instructions and code, so a bad Skill can make Claude do things that don’t match what it claims to do. Anthropic recommends using only Skills you wrote yourself or that Anthropic provides. In the worst case, a bad Skill could lead to data leaks or unauthorized access to your systems.
The official advice is to be as careful as you would installing software. If you really must use a Skill from an unknown source, check every file in it first. To learn more about the risks, see Security Risks of AI Agents.
Further reading
- The Complete AI Agent Guide
- Skills, Sub-agents, and Cron: Three Automation Mechanisms
- What Is MCP: covers how Claude connects to outside tools, a different topic from Skills
Cover image source: Anthropic (claude.com/blog/skills)