Products, laws and numbers here change fast. Checked October 2026.
In plain words
An AI agent is clever in general but knows nothing about how your team does things: your report format, your brand colours, the steps to file an expense. You could paste all of that into every chat, but it would crowd out the actual work.
A skill packages that know-how into a folder: a main instruction file, plus any templates, examples or small programs it needs. The agent sees only a one-line description of each installed skill. When a task matches, it opens that skill and follows it.
So an agent can carry dozens of skills, such as making slide decks, filling PDF forms or following your company's writing style, without being slowed down by all of them at once.
Everyday example
A cook with a shelf of recipe folders
A good home cook does not memorise every recipe. They keep a shelf of labelled folders: "Diwali sweets", "Kids' tiffin", "Guests for dinner". Most days they glance at the labels only. When guests are coming, they pull out that one folder and follow it, using the shopping list tucked inside. Skills are those labelled folders, and the agent is the cook.
Where the example breaks: A cook can tell when a recipe looks wrong. An agent tends to follow a skill's steps as written, so a badly written or malicious skill can lead it astray.
How it actually works
Anthropic introduced Agent Skills in October 2025 and published the format as an open standard that December.
A folder with SKILL.md
Each skill is a folder containing a file called
SKILL.md. It starts with a short header giving anameand adescription, followed by instructions. It can bundle scripts, reference files and templates.Load only the labels
At startup the agent reads just the name and description of every installed skill into its system prompt. That costs very little of the context window.
Open when relevant
When a request matches a description, the agent reads the full
SKILL.md. If that points to more files or scripts, it opens or runs them only as needed. The standard calls this "progressive disclosure".
- Introduced
- Anthropic, 16 October 2025, for Claude. Open standard published 18 December 2025, now at agentskills.io
- Adopted by
- As of October 2026, OpenAI's ChatGPT and Codex and Google's Gemini CLI document support for the open standard, among other agent products
- Example
- Anthropic's PDF skill lets Claude fill in PDF forms, which it could read but not easily edit before
- Safety
- Skills can contain code. Anthropic advises "installing skills only from trusted sources" and checking others before use
Questions people ask
What is a SKILL.md file?
It is the main file of an agent skill. A short header gives the skill's name and a description of when to use it, and the rest is plain instructions for the agent. Other files in the same folder, like scripts or templates, can be referenced from it.
What is the difference between agent skills and MCP?
MCP connects an agent to outside apps and data, such as your calendar or a database. A skill gives the agent know-how: steps and rules for a type of task. MCP is about access, skills are about method, and they work well together.
Are agent skills only for Claude?
No. Anthropic started them for Claude, then released the format as an open standard in December 2025. As of October 2026, OpenAI's ChatGPT and Codex and Google's Gemini CLI say they support it, so one skill folder can work across several agents.
How is a skill different from a system prompt?
A system prompt is read in full at the start of every conversation. A skill is loaded only when needed, apart from its one-line description. That lets an agent hold many skills without filling its memory with all of them.
Can I write my own agent skill?
Yes. A skill is just a folder with a SKILL.md file written in plain language, plus optional files. Describe clearly when it should be used, give step-by-step instructions, and test it on a few real tasks.