What an agent skill is
A skill is a folder that holds a SKILL.md file and, where needed, supporting files. The top of SKILL.md carries metadata — at minimum a name and a description — followed by instructions for the AI agent: how to do the task, what to check, when to run the scripts.
size-advisor/
├── SKILL.md # name, description, instructions
├── references/ # brand size charts
└── scripts/ # size conversion between systems
Anthropic introduced Agent Skills on 16 October 2025 in the Claude apps, Claude Code and the API. On 18 December 2025 the format was published as an open standard, with the specification at agentskills.io. As of September 2026, the list of compatible clients on the standard’s site includes ChatGPT and Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and other agents.
How an agent loads skills
Skills are loaded in stages (progressive disclosure):
- Discovery. At startup the agent reads only the name and description of each skill.
- Activation. When a task matches a description, the whole SKILL.md is loaded into context.
- Execution. The agent follows the instructions and runs scripts or opens reference files as needed.
That way many skills can be on hand while the context window is spent only on the one in use.
Skill vs MCP tool vs prompt
| Prompt | MCP tool | Agent skill | |
|---|---|---|---|
| What it is | Instruction text | Access to an external system | Instructions, scripts, materials |
| When it is in context | Always | Description always, result after the call | Description always, contents on demand |
| Question it answers | How to behave | What the agent can do | How to do the task |
| Retail example | “Answer only from the catalogue” | Order status from the order system | Returns procedure with rules and deadlines |
Important: Alexa Skills are a different concept. They are apps for Amazon’s voice assistant that the user launches; Agent Skills are read by the model.
Retail examples
- Size advice from a size chart. The skill lists which measurements to ask for, holds the brand size charts and a size-conversion script, so the answer is the same whichever agent asks.
- Processing a return. Rules, deadlines, required data and steps; the agent files the request through the order-system tool, and a person confirms the final step.
- Product descriptions to brand guidelines. Tone, structure, banned phrasing and a template for the agent that prepares product pages for upload to the feed.
In practice: keep skills in git, write a precise description — it decides when the skill fires — and read third-party skills before installing them: a hidden instruction inside works like a prompt injection.