What Are AI Skills and How to Use Them (2026 Guide)

An AI skill is a small folder that teaches an AI agent how to do one job well — and then loads itself only when that job actually comes up. Inside that folder is a single Markdown file (called SKILL.md) with the instructions, plus any extra files the task needs: reference docs, templates, even runnable scripts.

Anthropic calls them Agent Skills. They are the missing layer between a one-off prompt you type and a full custom app you have to build. You write the expertise once, and every future conversation can use it without you re-explaining anything.

We run a Skills Marketplace here at PromptsRush, so we read and test a lot of SKILL.md files. This is the plain-English version of what actually matters: what a skill is, how the format works, where you can use one, and how to put your first skill to work today.

What an AI skill actually is

Strip away the jargon and a skill is three things bundled in one folder:

• Instructions — a set of rules and steps written in plain Markdown that tell the agent how to do the task the way you want it done.
• Context — optional reference files (a style guide, an API cheat sheet, a checklist) the agent pulls in when it needs the detail.
• Tools — optional scripts the agent can actually run, so the skill can do things, not just describe them.

The agent does not read all of this all the time. It reads a one-line summary first, decides whether the skill is relevant, and only then loads the rest. That on-demand behaviour is the whole trick, and we will get to it below.

A useful mental model: a prompt is something you say; a skill is something your agent knows. Once a skill is installed, the knowledge is just there.

Skills vs prompts vs custom instructions vs MCP

"Skill" gets used loosely, so here is where it sits next to the things it is most often confused with.

The short read: prompts are for one-off asks, custom instructions hold your standing preferences, MCP connects your agent to live data and tools, and skills package a repeatable procedure the agent can run again and again. They stack — a single skill can call MCP tools and respect your custom instructions at the same time.

How skills work under the hood

Every skill starts with a SKILL.md file. The top of that file is a small block of YAML "frontmatter" with two fields that do the heavy lifting:

---
name: brand-voice
description: Rewrites copy in the PromptsRush brand voice. Use when the user asks to draft, edit, or polish marketing or blog copy.
---

# Brand voice

When rewriting copy:
1. Lead with the conclusion, then justify it.
2. Cut filler like "in today's fast-paced world".
3. Prefer short sentences in the opening paragraph.

Here is the part people miss: the description is a trigger, not a label. Before the agent ever opens the skill, it sees only the name and description. If the description clearly states what the skill does and when to use it, the agent loads it at the right moment. If it is vague, the skill sits there unused.

Progressive disclosure: why skills do not bloat your context

Skills are built around a principle called progressive disclosure, and it is the reason you can have dozens installed without slowing anything down:

• Level 1 — always loaded: just the name and description (a few words each). Cheap.
• Level 2 — loaded when relevant: the full body of SKILL.md, once the agent decides the skill applies.
• Level 3 — loaded on demand: any extra files the body points to — a long reference, a script, a template — pulled in only at the moment they are needed.

So a skill can be large on disk and still cost almost nothing until it is actually used. That is a real advantage over stuffing everything into one giant system prompt.

What is inside a skill folder

A simple skill is one file. A richer one looks like this:

my-skill/
|-- SKILL.md          (required - instructions + name/description)
|-- scripts/
|   `-- convert.py    (optional - code the agent can run)
|-- references/
|   `-- api-guide.md  (optional - loaded only when needed)
`-- assets/
    `-- template.docx (optional - files the skill works with)

Because it is just Markdown and ordinary files, a skill is portable. You can zip it, drop it in a Git repo, or share it — which is exactly how the skills in our marketplace travel.

Where you can use skills today

Skills are not tied to one surface. The same folder works across the places you already use Claude:

• The Claude apps (web and desktop) — turn on Skills in settings and Claude can use them in normal chats.
• Claude Code — drop a skill into your project's .claude/skills/ directory and it is available to your coding agent. This is where the developer-focused skills shine.
• The Claude Developer Platform and Agent SDK — attach skills to your own agents and apps through the API, so your product gets the same capability.
• Other agents — because the format is plain Markdown, tools like Cursor and other coding agents can read the same skill folders. The skills in our marketplace list Claude, Cursor and "other" for exactly this reason.

Want to see one in the wild? Open the GSAP AI Skills page — GreenSock's official skill that teaches a coding agent to write correct, performant animations. The detail page shows the SKILL.md, the file tree, and a one-click download.

How to use a skill: step by step

Using an existing skill takes about two minutes.

1. Find a skill for the job. Browse the PromptsRush Skills Marketplace by platform, category, or difficulty, and open the one that fits.
2. Get the files. Hit Download to grab the skill as a zip, or copy its SKILL.md straight from the page. Every skill is backed by a public GitHub repo, so you can inspect exactly what you are getting first.
3. Install it where you work. For Claude Code, unzip it into .claude/skills/. For the Claude app, add it through the Skills settings. For your own agent, attach it through the API.
4. Just ask. You do not "call" a skill by name — you describe your task normally ("animate this hero section", "add a Stripe checkout"), and the agent loads the matching skill on its own.
5. Check the result. Good skills make the agent more correct and more consistent. If nothing changed, the skill's description probably did not match your ask — more on that in the build guide.

Real skills you can try right now

Abstract explanations only go so far. These are free, open-source skills you can install today, and each one is a clean example of the format:

• GSAP AI Skills — GreenSock's official skill set. It stops coding agents from hallucinating GSAP syntax and gets you smooth, performant animations on the first try.
• Next.js, React and Vercel Skills — Vercel's official skills that teach an agent to build and ship Next.js and React apps the right way, from routing to deployment.
• Stripe AI Skills — Stripe's official skills for building secure, correct payment integrations instead of the broken boilerplate agents usually produce.

Notice the pattern: every one of these is a company packaging its own hard-won expertise so an agent stops getting it wrong. That is the sweet spot for skills.

Skills vs MCP: when to use which

This is the most common question we get, so here is the clean line:

• Use MCP when the agent needs to reach something live — your database, a SaaS API, today's tickets. MCP is a connection.
• Use a skill when the agent needs to know a procedure — how your team writes commits, how to assemble a report, how to use a tricky library. A skill is knowledge.

They are not rivals. A well-built skill often says "fetch the data with this MCP tool, then format it like this." Skills are the playbook; MCP is the phone line.

Why skills are worth the effort

• Reusable. Write the expertise once; use it in every future chat. No more pasting the same three paragraphs of context.
• Consistent. Everyone on the team gets the same behaviour, because everyone's agent reads the same skill.
• Token-efficient. Progressive disclosure means an unused skill costs almost nothing.
• Portable. Plain Markdown moves between Claude, Claude Code, your own agents, and other tools.
• Composable. Agents can combine several skills in one task without you orchestrating it.

The limitations (and the one real risk)

Skills are powerful, which means they deserve a moment of caution:

• They can run code. A skill can bundle scripts the agent executes. That is the feature — and the risk. Only install skills from sources you trust, and read the SKILL.md before you run it. Every skill in our marketplace links to its public GitHub source so you can audit it.
• A vague description equals a dead skill. If the agent cannot tell when to use it, it will not. The description does more work than the instructions.
• They are not magic. A skill makes a capable model more reliable at a task; it will not make a task possible that the model fundamentally cannot do.

How to get started

Two paths from here, depending on what you want:

1. Use skills. Browse the Skills Marketplace, install one that matches something you do often, and let your agent pick it up automatically.
2. Build your own. If you have a task you re-explain to Claude over and over, that is a skill waiting to happen. Our step-by-step walkthrough — How to Create Your First Claude Skill — takes you from an empty folder to a working, shareable skill.

And if you build something good, you can submit it to the marketplace so other people can use it too.

Pro tip: the fastest way to learn the format is to download one real skill, read its SKILL.md, and copy the structure. Start from GSAP AI Skills or the Vercel skills.

Keep learning

Skills pair well with the rest of your Claude workflow. If you are coding with Claude, see how developers are using Claude 4.8 for vibe coding and our Next.js prompts for Claude. For everyday firepower, the 100 best Claude prompts for power users is a good next stop. Or jump straight into the Skills Marketplace and the wider PromptsRush marketplace.