In one sentence: a Skill is an “expertise pack” you install into an AI agent — a folder containing an instruction manual and tools that teaches the AI how a specific task should be done professionally.
Analogy: onboarding training for a new hire
An AI agent is like a smart employee with no work experience. Ask it to “write a weekly report” and it can — just not professionally. What a Skill does is hand that newcomer a job manual + toolbox: install a “Weekly Report Skill” and it starts writing in your company’s format, terminology, and the structure your manager prefers. Install a “Design Skill” and it produces professional UI. Different Skills are training packs for different roles.
How it works: progressive disclosure, no wasted context
At its core, a Skill is a standard folder structure built around a SKILL.md file:
---
name: my-skill-name
description: One line on what this skill does and when to use it
---
# Skill instructions (main body)
[Detailed steps, rules, examples...]The AI only reads a skill’s content when it is relevant to the task — this is called Progressive Disclosure: it first reads the name and description to decide whether to use the skill, then reads the body and attachments when needed. This keeps the context window from overflowing, so the agent stays fast and reliable.
How Skills became an open standard
The format was initiated by Anthropic and released as an open standard on December 18, 2025 (agentskills.io). It has since been adopted by roughly 40 mainstream AI clients, with nearly every major AI tool gradually joining in.
A single standard means: write one Skill, and it works in every AI tool that supports the standard.
Skills vs. MCP: what’s the difference?
| Skill | MCP | |
|---|---|---|
| Essence | A manual that teaches AI “how to do it” | A connector that lets AI “reach what’s outside” |
| Analogy | An onboarding handbook | A USB-C port (plug in external tools) |
| Contents | Text instructions + scripts + reference files | Code servers providing tool/resource access |
| Who it’s for | Everyday users can install and use them | Usually requires configuration; more developer-oriented |
How they relate: a Skill can call MCP tools from its scripts/ directory — which is why they are often used together.
Summary
- Skill = expertise pack / training manual; MCP = tool connector. They complement each other.
- One Skill = one folder with a
SKILL.md, loaded on demand. - It is an open standard supported by ~40 mainstream AI clients.
- Want ready-made Skills? Browse our curated list. Want to write your own? See the tutorial.