Skip to content
InfratoApp
InfratoApp / Blog / what-claude-skills-are-and-where-they-live

Claude skills: what they are and where to find them

A Claude skill is a folder with a SKILL.md file: a name, a description and instructions a model loads. Here is what that means and where skills come from.

INFRATOAPP · PUBLISHED 18 OCTOBER 2026 · 4 MIN READ · 3 SOURCES · PRICES LAST CHECKED 18H AGO

What a Claude skill is

A Claude skill is a folder that holds a file called SKILL.md. The file opens with a short block that gives the skill a name and a description, and the rest is plain instructions for a model to follow. The model reads the description to decide whether the skill fits the task in front of it, and loads the instructions when it does. A skill is text, so by itself it reaches no service and holds no credential.

The directory calls these agent skills, because the file format is what defines them and not the assistant that happens to load one. Searching for Claude skills and searching for agent skills lead to the same kind of thing, and the skills directory lists them under the second name.

What is inside the folder

The front block is flat: a name and a description, and sometimes a licence. It is written as simple key and value lines, which is why a skill can be read by a person in a few seconds. The description is the part that matters most. It is the sentence the model weighs against everything else it could do, so a vague one fires at the wrong time and a precise one fires when it should.

Below the block sits the body: the steps, the checks, the formats and the examples the skill wants followed. Some skills keep everything in that one file. Others ship extra files beside it, such as reference notes or scripts, and the body points to them. Whatever a skill can do beyond text lives in those extra files, so they are the part worth opening.

Where skills come from

Skills are published in repositories, and one repository often holds many of them. Anthropic publishes its own in a repository of that name, companies publish skills about their own products, and individuals publish whatever they have written for themselves. That is why a directory of skills is really a directory of repositories, read one file at a time.

The directory holds 3,022 skills on record. It counts a skill only when it resolves to a SKILL.md that was actually read, with the address of that file beside it, and it marks the ones a vendor published about its own product. Other indexes list far more names. The difference is mostly entries that point at nothing we could open, and the directory leaves those out rather than pad the figure.

Where you find a skill matters less than whether you can open the file behind it. A listing with no file to read is a claim and not a skill.

How a skill differs from a prompt and from an MCP server

A prompt is something you type each time. A skill is the same kind of guidance saved once, so a model can pick it up when the task calls for it without you pasting it in. The saving is in repetition: a job you explain the same way every week is a candidate for a skill.

An MCP server is a different tool for a different problem. It gives a model a way to reach a system it cannot see, and it usually needs a credential to do so. A skill gives a model instructions for work it can already do. The two are often used together, and choosing between them is a question about access first and instructions second.

What a skill cannot do

A skill cannot connect to anything. It cannot read your files, call an API or run a command unless the assistant you load it into already has that ability and chooses to use it. What it contributes is instructions, and instructions are advice: a model can follow them closely, loosely or not at all, and no file can force the choice.

A skill can also be wrong. It is written by one person for one setup, it can go out of date as the tools it mentions change, and nothing checks it. The description decides when it is used, so a skill that fires too often is as much a problem as one that never fires. Treat a skill the way you would treat a colleague's checklist: useful, and yours to edit.

Starting with one

Pick a task you do often and already know how to do well. Find a skill that claims to cover it, open its SKILL.md, and read the description and the body before you install anything. If the instructions describe a process you recognise, load the skill the way your assistant says to, and run the task once to see whether it helps.

The longer checklist, covering the publisher, the licence and how to test, is in choosing agent skills from a directory. Start small. One skill whose description is sharp will teach you more about the format than a dozen you never read.

What we could not answer

  • ⚑The directory does not test whether a skill works, and it holds no usage data, so a listing says a skill exists and where it lives, not that it is good.
  • ⚑It lists skills by their file format, so it cannot say which assistants load which skills or how any assistant decides to use one.
infratoapp, “Claude skills: what they are and where to find them”, read 27 October 2026. agent skills in the directory: 3,022. https://www.infratoapp.com/blog/what-claude-skills-are-and-where-they-live

Questions people ask about this

What is a Claude skill?

A folder that holds a SKILL.md file: a name and a description at the top, then plain instructions. A model reads the description to decide whether the skill fits the task, and loads the instructions when it does. The file is text, so a skill on its own reaches no service and holds no credential.

Do Claude skills need an API key?

Not for the skill itself, which is instructions and nothing else. A skill can tell a model to use a tool you have already connected, or ship a script beside the instructions, and those are where a key or a permission could come in. Read the extra files before you load a skill.

Where can I find Claude skills?

In repositories that hold SKILL.md files. The skills directory on this site lists the ones it has read, each with the address of its file, the publisher and the licence where one is stated. You can also open a publisher repository directly and read the skills it contains.

These render as FAQPage structured data · answers are filled from the same live figures

Read next

Agent skills or MCP servers: which do you need

Use an MCP server when the model must reach a system it cannot see, and a skill when it already has the access but needs to know how to do the task.

Choosing agent skills from a directory

Pick an agent skill by reading its SKILL.md, checking who publishes it and where it lives, and testing it on one task you already know how to do.

What an MCP server is, in plain terms

An MCP server is a small program that lets an AI assistant use tools and data outside the chat, such as files, a database or a ticket tracker.