What Are Skills and How to Use Them

Ask an AI to draft a weekly customer update, and the first one comes out fine. A week later, ask again, and the second draft lands in the same shape. Same sections, same tone, the disclaimer in the same place. Nothing was re-explained between the two.

That isn't memory at work. Memory holds facts about a person: their name, their preferences, their birthday. The reason the second update matched the first is a skill, a small file the AI knew to follow when that kind of work came up.

Skills are how an AI learns to do a particular job the same way every time. Anthropic developed the format inside Claude and released it as an open standard at agentskills.io, and most of the agent products people are already using have started supporting it. The spec fits on a page, the format belongs to nobody, and if "skills" has been showing up in your feed without quite settling into a meaning, this is the explainer.

A skill is a folder

A skill is a folder containing a file called SKILL.md, and that file is the AI's instructions for doing one specific job. The folder can include extra things if the work needs them, like a script or a template or a reference doc, but plenty of useful skills are one folder with one file.

Here's a real one, simplified:

---
name: weekly-client-update
description: Drafts a weekly client update with what got done,
  what's blocked, and next steps. Use when the user asks for a
  weekly update, status report, or end-of-week recap for a client.
---

When the user asks for a weekly client update:

1. Pull what got finished this week from the project notes.
2. List anything blocked, with the reason.
3. List the top three things planned for next week.
4. Format as a short email, friendly but professional.
5. Sign off with the user's name.

Two things are required at the top: a name and a description. Everything below is the AI's playbook for the job, written the way you'd brief a new hire on day one. The description is the most important line in the whole file, and it's worth understanding why before writing one of these.

How the AI knows when to use a skill

When a conversation starts, the AI loads the names and descriptions of every skill available to it, and the full instructions for each one stay on disk for the moment. When a request comes in, the AI checks whether any of those descriptions match. If one does, it pulls in the full instructions for that skill and follows them, while the rest stay in the background.

That progressive loading is the entire reason this works. An AI can only hold so much in context at once, and if every skill's full instructions loaded at the start of every conversation, the workspace would be cluttered before anyone said a word. Loading what the moment calls for keeps the AI sharp.

It's also why the description matters more than the file name or anything else in the body. "Helps with reports" is too vague to fire reliably, because the AI has nothing concrete to match against. Something like "Drafts a weekly client update with what got done, what's blocked, and next steps; use when the user asks for a weekly update, status report, or end-of-week recap" picks up the moment someone types "give me my Friday update," without anyone needing to name the skill at all.

What skills are good for

The use cases that come up most often are jobs done with AI more than twice that need to look the same every time: weekly client updates, listing descriptions for a real estate agent, lesson plans for a teacher, pre-publish checklists for a podcast, refund-request responses for an online store. Most of that work currently lives in someone's head, in a Notion doc nobody opens, or in a Slack thread from three months ago, and a skill is what turns it into the AI's working knowledge.

Skills are also the right home for procedures that depend on details only one person on a team knows: the exact order operations get done in, the phrase that has to appear in the closing line, the unit a number always reports in. When that knowledge lives in a skill, it stops walking out the door when the person who knows it leaves.

And because the format is open, a skill written today still runs in a different agent product a year from now. The folder comes with you.

What skills aren't

Most of the confusion around skills comes from collapsing them into something nearby. The clearest way to understand the format is by what it isn't.

Not a prompt

A prompt is a single message someone sends to an AI in a chat. A skill is a set of instructions the AI loads on its own when a request matches its description, with no command to remember and no menu to open.

Not memory

Memory holds facts about a person: their name, their preferences, the dog. Skills hold procedures: how to draft the update, how to format the recap, how to handle a refund. Useful AI setups generally have both, and the memory explainer covers the memory side in depth.

Not a connector

A connector lets the AI reach Gmail or a calendar or a database. A skill tells the AI how to use those tools well once it has them. Connectors and skills sit at different layers, and they work together; the post on skills vs MCP gets into how.

Not magic

The instructions inside are written in regular English, the same way you'd describe the work to a new hire. A vague skill produces vague output, no matter how capable the model underneath happens to be.

Where the standard came from

Anthropic shipped skills inside Claude, then published the format publicly so anyone could use it. The list of agent products that have adopted it has grown steadily and now includes Claude Code, Cursor, GitHub Copilot, OpenAI Codex, OpenHands, Goose, and a long tail of others, all reading the same SKILL.md format.

The reason this small idea is spreading while bigger ones sit in committee is that there's not much to it. The spec fits on a page, there's no runtime to install, no service to wire up, and no SDK to learn. The whole format is shorter than this article.

How skills work on Zo

On Zo, skills live in the workspace as regular folders. Not behind a login on someone else's server and not in a vendor's database, but in a folder called Skills/ inside your computer, where you can open them, read them, edit them, and share them.

Every Zo ships with the runtime built in. Install a skill from the hub or write a new one in chat, and the folder lands in the workspace. Zo reads from it. You can change it whenever you want, and if you ever move to a different agent product, the same folder will keep working there.

A few skills already running in Zo workspaces today:

None of those are products you can buy. They're folders, and anyone with a text editor can open them.

Where to go next

The practical follow-up is How to Make a Skill That Actually Works, which walks through writing, testing, and maintaining a skill in one conversation. For where skills sit alongside the other AI standard people are hearing about, Understanding the Differences Between Agent Skills and MCP Servers covers how the two fit together. To browse real skills and pick one up, the Zo Skills hub is the fastest way in.

Frequently asked questions

A skill is a folder with an instructions file. Once it's there, an AI does the same job the same way every time. Get started with Zo Computer.