Claude Skills: what they are and how to install

Claude Skills: what they are and how to install

Claude Skills are folders of instructions that Claude reads only when a task actually needs them. Instead of pasting the same guidance into every conversation, you package it once and Claude pulls out the right skill on its own. Anthropic launched the feature in October 2025 and it works in Claude Code, in the Claude app and through Anthropic's API. Here is what a skill is, how it differs from MCP and CLAUDE.md, how to install and build your own, and what to watch out for before you download one from the internet.

If you want the basics of the assistant itself first, they are in our guide to Claude AI.

What is a skill in Claude?

A skill is a folder containing a file named SKILL.md. The file has two parts: a frontmatter block that describes the name and purpose, followed by plain markdown with the actual instructions. The only required field is description — the text Claude uses to decide when the skill should be used.

The folder can also hold reference files, templates and scripts. Think of a colleague handed a binder: the cover says what it is about, and the colleague only opens it when the task comes up. A skill works the same way, which is why you can have many installed without anything getting heavy.

How do Claude Skills work?

The trick is that Claude does not read everything at once. The pattern is called progressive disclosure and has three steps:

1. Name and description for every installed skill are always in Claude's context. That is a line or two per skill.
2. The full SKILL.md is loaded only when your request matches the description.
3. Extra files and scripts in the folder are read or run only if the instructions point to them.

So you can have hundreds of skills without paying tokens until one is used. That is the big difference from stuffing everything into one long prompt.

Skills, MCP and CLAUDE.md — what is the difference?

The three are often confused, but they solve different problems:

• CLAUDE.md is the project's permanent memory note. It is loaded every time, so it should be short and always relevant.
• A skill is a way of working that is loaded only when needed. It suits procedures: "how we write a quote", "how we review a contract".
• MCP (Model Context Protocol) connects Claude to external tools and data, such as a CRM or a database. More in our guide to AI agents.

Rule of thumb: MCP gives Claude hands, a skill gives Claude knowledge of how the job should be done. They are often combined — a skill can describe how Claude should use an MCP tool your way.

How a skill gets triggered

By default both you and Claude can start a skill. You type /skill-name to run it directly, or Claude notices that your request matches the description and loads it automatically. Two frontmatter flags control this:

• disable-model-invocation: true — only you can trigger the skill, never Claude. Good for anything with side effects, like a deploy or a send-out.
• user-invocable: false — the reverse: only Claude can use it, and it does not show up as a command.

Because description decides whether Claude finds the skill, that line is worth the time. Write what the skill does and in which situations it should be used.

How do I install a Claude Skill?

Where you put the folder decides who gets access:

• Personal — ~/.claude/skills/<name>/SKILL.md, available in all your projects.
• Project — .claude/skills/<name>/SKILL.md, applies only to that repo and can be checked into git so the whole team gets it.
• Plugin — skills shipped with a plugin, namespaced as plugin-name:skill-name so they never collide with your own.
• In the Claude app — under Customize → Add you can add a skill without touching the terminal.

Claude Code picks up new and changed skills live during the session. No restart is needed as long as the folder existed when the session started.

How to create your own skill

Writing your first skill takes a few minutes:

1. Create the folder: mkdir -p ~/.claude/skills/my-name
2. Write a SKILL.md with frontmatter and instructions.
3. Test it by typing /my-name, or ask a question that matches the description and check that Claude finds it on its own.

You can also ask Claude to write the skill for you in the app. Start small: a skill that solves one recurring task works better than one that tries to cover everything.

Examples of Claude Skills

A few uses that show what the format suits:

• Document files. Anthropic has ready-made skills for creating and editing PDF, Word, Excel and PowerPoint files.
• Brand voice. A skill with your tone, banned words and sample texts, so every draft sounds like your company.
• Quote template. Structure, standard terms and price lines that Claude follows every time.
• Code conventions. How you name, test and review code in a given project.

We use a skill ourselves to write blog articles to a fixed style guide. It keeps the texts in the same voice no matter who asks for them.

Skills on skills.sh and GitHub

You do not have to build everything yourself. skills.sh is an open marketplace for skills that work with Claude Code, Cursor, Codex and several other agents. A skill installs with one command:

npx skills add <owner/repo>

It is the same idea as npm install, but for agent instructions instead of code. There are also large collections on GitHub with hundreds of skills — which is exactly why the next section is the most important in this guide.

Are Claude Skills safe?

A skill you install is text that Claude trusts as instructions — not code reviewed by a compiler or run in a sandbox. If a SKILL.md contains a hidden prompt injection, instructions asking Claude to do something other than what you asked, like reading out secret keys or sending data to an external service, no automatic boundary stops it. Skills with an allowed-tools field can also give Claude the right to run commands without you approving each call, which makes a malicious skill more dangerous than a malicious snippet of text.

Before you install a skill from skills.sh or another repo:

• Read SKILL.md yourself, especially if it has an allowed-tools field.
• Check which tools it asks for. A skill that only reads files is a very different risk from one that wants to run arbitrary commands.
• Use the audits on skills.sh if you are unsure.
• Prefer known publishers and open source you can actually read over an anonymous single file.

The same common sense that applies to npm install of a package you have never heard of applies here. The difference is that the consequence is an AI assistant acting against you.

Are Claude Skills worth using?

Yes, for recurring work where the result should come out the same every time. A skill is rarely worth the effort for a one-off task. Some in the community say skills are "just prompts in a folder", and technically that is right: they do not give Claude any new abilities. The value is in the knowledge being packaged, shared and loaded only when needed.

If you want help building skills for your company's workflows, we are happy to look at it together — read more about our AI consulting.

Summary

Claude Skills make it easy to package knowledge that Claude would otherwise forget between sessions, from internal code conventions to whole workflows. Build your own for what is specific to you, and use the community for what is already solved — but treat every skill you did not write yourself as code you would code-review, not as a harmless text file.