A Claude skill is a folder with a file called SKILL.md that teaches Claude how to do one job. The file starts with a short header giving the skill a name and a description of when to use it, followed by plain instructions. Claude keeps only the name and description in view, and reads the full instructions when a request matches. You add a skill once, and Claude can follow the same instructions each time, without you explaining them again. A skill doesn’t guarantee a right answer, so test it on real examples and check what matters.

What a skill is

Anthropic calls them Agent Skills: reusable packages of instructions, and optionally scripts and reference files, that Claude loads when they are relevant. The difference from a prompt is when it loads. A prompt is for one conversation. A skill sits in the background, and only its name and description take up room in Claude’s context until a request calls for it. That is why you can have many skills without slowing anything down, and why the description matters so much: it is what Claude matches your request against.

Anthropic’s overview describes three layers, each loaded only when needed:

What goes in a SKILL.md

The file is Markdown with a YAML header between two lines of three dashes. Two fields are required, with these rules:

FieldRulesWhat it is for
nameUp to 64 characters: lowercase letters, numbers and hyphens. It can’t contain the words “anthropic” or “claude”.What the skill is called.
descriptionNot empty, up to 1,024 characters, no XML tags.What the skill does and when to use it. Claude picks a skill by this line.

After the header comes the body: the instructions, in whatever headings make them clear. Anthropic recommends keeping the body under 500 lines and moving detail into separate files that SKILL.md links to directly, one level deep. A skill that is only a SKILL.md is complete; scripts and extra files are optional.

A small example

An invented skill for one job: turning meeting notes into action items.

---
name: meeting-action-items
description: "Turns meeting notes or a transcript into a list of action items, each with an owner and a due date. Use when the user shares meeting notes and asks for action items, next steps, follow-ups, who owes what, or what was agreed."
---

# Meeting action items

Turns meeting notes into a short list of who does what, by when.

## Steps
1. Read the notes or transcript in full. If none were shared, ask for them before doing anything else.
2. Pull out every commitment: something a person agreed to do. Skip ideas nobody took on.
3. For each one, give the owner and the due date exactly as the notes say. If either is missing, write "Not stated". Never guess.
4. Group the items by owner, people in the order they first spoke.
5. End with a section called Open questions, listing anything raised but not decided.

## Done when
- Every item can be traced to a line in the notes.
- No owner or date was invented.
- The list fits on one screen. If it doesn't, say how many items there are and ask whether to shorten it.

## Don't
- Don't email or message anyone. Show the list and stop.
An example SKILL.md. Save it in a folder named meeting-action-items.

How to write a good one

One job per skill

“Meeting action items” is a skill. “Help with meetings” is not. If you find yourself writing “or” in the description, you probably have two skills. Small skills are easier to describe, easier for Claude to pick, and easier to fix.

A description that says when

Write it in the third person (“Turns meeting notes into…”, not “I can help you…”), say what it makes, then say when to use it, using the phrases people actually type: “next steps”, “follow-ups”, “who owes what”. Anthropic’s guidance is blunt about vague descriptions such as “Helps with documents”: Claude has nothing to match them against.

Steps in order, each one action

Number the steps. Say where the input comes from and what to do if it is missing. Say what to keep and what to leave out. Where getting it wrong would be costly, be exact; where judgment is fine, leave room. Skip explanations Claude doesn’t need: it already knows what a meeting is.

Say what done looks like

A short list of checks to run before handing the result over does more than any amount of “be careful”: every item traceable to the source, nothing invented, the length you want. This is also where you say what the skill must never do, such as sending anything on its own.

Test it on real work

Run it on three or four real cases, including a messy one. When it gets something wrong, change the instruction that allowed it, not the wording of your request. Anthropic suggests working with Claude itself: do the job once together, then ask Claude to turn what you explained into a skill, and trim what it writes.

Only add skills you trust. A skill is instructions, and sometimes code, that Claude follows. Anthropic advises using skills you wrote yourself or got from a trusted source, and reading every file in one you didn’t.

Where to add a skill

A skill added in one place is not available in the others, so add it wherever you want to use it.

Menu names move from time to time. If a step looks different, look for Skills in Claude’s settings.

Can I use a SKILL.md in ChatGPT?

Whether ChatGPT can use a skill file directly depends on your plan and workspace settings; OpenAI’s help center has the current details, so check whether yours offers them. Where it doesn’t, the file still works as plain instructions: paste its steps into a chat, or keep them somewhere you can reuse, such as your custom instructions if the job is one you do in most chats. You lose the automatic part, where the assistant picks the skill up by itself, but you keep the brief.

Turning a job you repeat into a skill

  1. Pick a job you have explained to AI more than twice.
  2. Write down what starts it, what it works from, the steps, and what you hand back.
  3. Write the description last, from the words you actually use when you ask for it.
  4. Add what done looks like, and what it must never do.
  5. Run it on real work, and fix the instruction behind each mistake.