How Do Claude Skills Work? A Simple Explainer

Claude skills work by storing a task's instructions in a folder that Claude loads only when it's needed, so it can carry many skills at once without slowing down or getting confused. When you ask for something that matches a skill, Claude opens that skill, follows its steps, and produces the result. This short guide explains how Claude skills work under the hood, in plain English, and why the design is so effective.

If you're new to the idea, a skill is a reusable set of instructions that teaches Claude to do a specific task your way.

Watch: skills are covered in full in our free Claude Cowork course.

The filing-cabinet idea

Picture a filing cabinet. Without skills, giving Claude lots of capabilities means cramming every guideline you have, for emails, reports, social posts, into one enormous prompt that sits in front of it all the time, causing confusion. With skills, Claude just knows which folders exist in the cabinet. Ask for a branded email and it opens the email folder, reads those instructions, and gets to work, while everything else stays in the drawer.

That's the whole mechanism in a sentence. The detail worth knowing is how the folder is structured.

The three layers of a skill

Every skill has three layers, and they load at different moments.

The first layer is the metadata: the skill's name and a short description. This is always loaded, so Claude always knows the skill exists and when it's relevant. It costs almost nothing, which is why you can install many skills without slowing anything down.

The second layer is the instructions: a longer file with the step-by-step workflow. Claude only reads this when the skill is actually triggered, so the detailed guidance enters its working memory at the moment it's needed, not before.

The third layer is the resources: supporting files like templates, fonts, logos, or saved scripts. These load only when a task calls for them. A skill might bundle a dozen files, but Claude opens just the one it needs.

Why this design matters: progressive disclosure

Loading information in stages, only as it's needed, is called progressive disclosure, and it's the reason skills are both powerful and cheap to run.

It means Claude isn't holding every instruction in memory at once, so it stays focused and doesn't mix up guidance from unrelated tasks. It means you can keep dozens of skills available without crowding the context window or burning through tokens. And it means a skill can include comprehensive reference material or scripts that only get read when relevant, so detail never costs you performance.

There's a knock-on benefit, too. Because a saved skill runs as a recorded process rather than something Claude works out from scratch each time, the results are faster and more consistent. The expensive trial-and-error happens once, while you build it.

What this means in practice

You don't need to manage any of this; Claude handles the loading. But understanding it changes how you build. It's why a good skill leads with a clear, specific description, that's the always-loaded layer Claude uses to decide when to reach for it. It's why bundling extra reference files into a skill is free rather than costly. And it's why skills scale: a whole team can carry a library of them, each loading only when its moment comes.

If you want to put this into practice, our guides to using Claude skills and creating your own take it step by step, and our explainer on skills versus agents clears up where skills fit alongside the agent running them.

A real skill, layer by layer

It's easier to see this with one of our own. Our 9x video-highlights skill turns a workshop recording into five LinkedIn clips. Watch the three layers do their jobs.

The metadata is a short description telling Claude this skill exists for "finding highlights and cutting clips from a 9x Live recording." That line, and only that line, is what Claude reads to decide the skill is relevant when we ask it to "repurpose this week's workshop."

The instructions then load: read the transcript, pick eight to ten candidate moments, hand each to a visual analyser, cut the best five, and append our branded outro. Claude only reads this once the skill is triggered, so it isn't carrying those steps around during unrelated work.

The resources load last and selectively. The skill bundles our fonts, our outro, and helper agents, but on any given run Claude opens only what that clip needs. It's the same principle as the rest of the skill: nothing is loaded until the moment it earns its place.

Why the description does the heavy lifting

If there's one part of a skill to get right, it's that always-loaded description. As our co-founder Alex puts it, "the description at the top matters most: it's the only part Cowork sees before a skill is triggered." Get it specific and Claude reaches for the right skill on a loose request like "sort out this week's clips." Leave it vague and two similar skills compete, and the wrong one wins.

This is also why skills are specific rather than generic. We don't keep a catch-all "video editor" skill; we keep a 9x Live highlights skill for that exact workflow, with the exact steps and assets it needs. The narrower the skill, the more reliably its description triggers it.

Skills, instructions, and projects

Skills are easy to confuse with two neighbouring ideas. A skill covers one specific task. Instructions are general guidance Claude reads before every task, good for your role, your tone, or rules like file-naming conventions, and you can set them globally or per folder with a claude.md file. Projects are a friendlier working folder that keeps its own files, instructions, and memory. Skills sit inside all of this: the instructions shape how Claude works in general, and skills handle the specific jobs you repeat.

Common questions about how skills work

Do lots of skills slow Claude down? No. Only the metadata is always loaded, at roughly a hundred tokens per skill, so you can install many without a performance cost.

Where are skills stored? As folders on the virtual machine Claude works in, each with a SKILL.md and any resources. You can download one as a file to share it.

Can a skill include scripts and big reference files? Yes, and they cost nothing until used, because they only load when a task calls for them.

Why this beats a giant prompt or a custom GPT

If you've tried to give a chatbot lots of capabilities, you've probably done it the old way: one enormous system prompt with every guideline crammed in, or a custom GPT loaded with instructions for every job. The problem is that all of it sits in the model's attention at once, so a request about a LinkedIn post gets coloured by your blog rules and your newsletter rules, and the output drifts. Skills solve exactly this. Because only the relevant skill's instructions load when triggered, Claude follows one clean set of guidance at a time. You get the breadth of many capabilities without the confusion of holding them all in view, which is the practical reason skills scale where a single mega-prompt falls over.

The takeaway

Claude skills work like a well-organised filing cabinet: Claude always knows which skills exist, but only opens the one a task needs, and only reads its detail and resources at the moment they're required. That staged loading is what makes skills fast, cheap, and scalable across a team.

If you'd like to build a library of skills around how your team actually works, our hands-on Claude Cowork training takes you from your first task to skills the whole team can run.