What is an Agent Skill?

An Agent Skill saves a useful way of working so an AI agent can follow it again.

Imagine you carefully explain a task to Claude or Codex. You give it the steps, show it what good work looks like and point out the mistakes to avoid. The result is useful.

A week later, you need the same job again. Without a Skill, you have to find the old conversation or rebuild those instructions from memory.

What does an Agent Skill do?

An Agent Skill gives an AI agent more than a request. It gives the agent the directions and supporting material it needs to do one kind of work properly.

Those directions might explain which steps to follow, what to check first, how to make a decision and what the final result should contain. The Skill can also carry examples, reference material, templates or scripts when the job needs them.

The AI still performs the work. The Skill saves the human thinking that guides it.

Why are Agent Skills useful?

Most useful AI work does not come from one clever sentence. It comes from giving the AI enough context, a sensible order of work and a clear idea of what a good result looks like.

Without a Skill, those instructions often end up in an old chat, one very long prompt or one person's memory. They are difficult to improve and annoying to repeat.

A Skill gives those instructions somewhere stable to live. You can use them again, improve the weak parts and share the better version with someone else.

This does not make the AI model itself smarter. It stops the thinking around the task from being disposable.

How does an Agent Skill work?

The basic flow is simple:

1. You ask an AI agent to do a job.
2. The AI product checks whether one of its available Skills matches your request.
3. If it finds a match, the agent reads the Skill's instructions.
4. It opens any supporting files or runs any allowed scripts it needs.
5. It completes the job using those directions.

The open Agent Skills format stores all of this in a folder:

my-skill/
├── SKILL.md
├── scripts/
├── references/
└── assets/

Only SKILL.md is required. It names the Skill, explains when it should be used and contains its main instructions. The other folders are optional places for code, detailed information, templates or files used in the result.

This is the technical shape of a Skill, not the reason it is useful. The folder is the container. The useful part is the way of working saved inside it.

What can an Agent Skill look like in practice?

Here's one example.

Disk Sweep is an Agent Skill for reviewing what is taking up space on a Mac. The problem is easy to recognise: you need more storage, but deleting the wrong file could create a much larger mess.

A broad request such as “help me clean my Mac” leaves most of the judgment to the AI. It might identify large folders without explaining whether they are safe to remove or how difficult they would be to recover.

Disk Sweep gives the agent clearer directions. It reviews the available storage information, sorts each item into a cleanup decision and reports the recovery cost and confidence behind that recommendation. The person can then decide what to remove, what to keep and what needs another look.

This is one way a Skill can help. The important part is not Mac storage or the exact result Disk Sweep returns. It is that useful decisions have been written down so the agent does not have to invent the approach each time.

When should you use an Agent Skill?

Use a Skill when the quality of the result depends on instructions you expect to need again.

A task may be a good fit when:

• you keep rewriting the same long prompt;
• the steps need to happen in a particular order;
• the work depends on rules, judgment or examples;
• the agent needs the same references or templates each time;
• another person should be able to use the same way of working; or
• a vague answer could cause a costly mistake.

You do not need a Skill for every request. If “summarise this paragraph” already gives you the result you need, a normal prompt is enough.

The useful question is not whether the task sounds important. It is whether there is a good way of doing the work that deserves to be saved.

What is an Agent Skill not?

• It is not the AI agent. The agent does the work. The Skill gives it directions to follow.
• It is not an app. Claude, Codex or another compatible product provides the place where the Skill runs.
• It is not a tool. A Skill can tell an agent when and how to use a tool, but the tool performs the action.
• It is more than a saved prompt. A Skill can keep its instructions together with the files and resources the work needs.
• It is not proof that the instructions are good. Saving a bad way of working only makes it easier to repeat.

How do you start using Agent Skills?

The easiest way to understand Agent Skills is to try one on a job you already recognise. Check what the Skill promises, what information it needs and what it will return. Then compare the result with the same task done through a fresh prompt.

If you want to create your own Skill, begin with one repeated job. Write down the result you want, the steps that produce it and the decisions that stop the work from going wrong. Test those instructions on real examples before adding more files.

A Skill can begin as one file. It should become more complex only when the work genuinely needs more support.