Claude Skills — what they are, how they work, and how to install them

Claude Skills — what they are, how they work, and how to install them

Claude Skills is Anthropic's way of letting Claude learn new, reusable ways of working — packages of instructions, and sometimes scripts, that Claude only loads when they're actually needed. Instead of pasting the same instructions into every conversation, or letting a CLAUDE.md file grow into a long list of procedures, you package the knowledge once and Claude pulls it in when a task matches. Here's how they work under the hood, how to build your own, and what to think about if you pull ready-made skills from the community.

What is a skill, really?

A skill is essentially a folder containing a file named SKILL.md. The file has two parts: a frontmatter block in YAML at the top (between --- lines) describing its 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 clever part is how little it costs to have many skills installed at once. Claude Code only loads each skill's name and description into the system prompt from the start. Only when a task actually matches a skill does the full SKILL.md file get loaded into context — a pattern often called progressive disclosure. A skill can also point to its own reference files, templates, or scripts in the same folder, which in turn only get loaded when needed. That means you can have hundreds of skills installed without it costing any tokens until one of them is actually used.

How does a skill get triggered?

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

• disable-model-invocation: true — only you can trigger the skill, never Claude on its own. Good for anything with side effects, like a deploy or sending a Slack message.
• user-invocable: false — the opposite: only Claude can use the skill, and it won't show up as its own command in the menu.

Where does a skill live?

Where you place the folder determines who gets access to it:

• Personal — ~/.claude/skills/<name>/SKILL.md, available across all your projects.
• Project — .claude/skills/<name>/SKILL.md, scoped to that repo and can be committed to git so the whole team gets it.
• Plugin — skills bundled with a plugin, namespaced as plugin-name:skill-name so they never collide with your own.

Building your first skill

Creating your own skill takes a couple of minutes:

1. Create the folder: mkdir -p ~/.claude/skills/my-skill
2. Write a SKILL.md file with frontmatter and instructions, and save it in the folder.
3. Claude Code picks up new and edited skills live during the session — no restart needed as long as the folder already existed when the session started.
4. Test it either by typing /my-skill directly, or by asking something that matches the description and watching Claude pick it up on its own.

Getting ready-made skills from skills.sh

You don't have to build everything yourself. skills.sh is an open marketplace for skills that works with Claude Code, Cursor, Codex, and several other agents. You install a skill with a single command:

npx skills add <owner/repo>

It's the same idea as npm install, just for agent instructions instead of code — which is exactly why you should think twice before running it.

Security: a skill is instructions Claude trusts

This is the part that matters. A skill you install becomes text that gets loaded into Claude's context as instructions — not code that gets reviewed by a compiler or run in a sandbox. If a SKILL.md file contains a hidden prompt injection — instructions telling Claude to do something other than what you asked, like leaking secret keys or sending data to an external service — there's no automatic boundary stopping it. Skills with an allowed-tools field can also grant Claude permission to run things like Bash commands without asking you each time, which makes a malicious skill far more dangerous than a malicious snippet of text.

Before installing a skill from skills.sh or another community repo:

• Read the SKILL.md yourself before running it, especially if it has an allowed-tools field.
• Check what tools it asks for. A skill that only needs to read files is a very different risk than one that wants to run arbitrary shell commands.
• Use skills.sh's audits section if you're unsure — the platform lists reviews of popular skills.
• Prefer known sources, or repos with many installs and open source you can actually read, over a single anonymous file.

The same common sense that applies to npm install-ing a package you've never heard of applies here — except the consequence of trusting the wrong instructions is your AI assistant acting against you, not just a broken build.

Summary

Claude Skills makes it easy to package knowledge Claude would otherwise forget between sessions — everything from internal coding conventions to entire workflows. Build your own for what's specific to your team, and lean on the skills.sh community for what's already solved — but treat every skill you didn't write yourself as code you'd code-review, not as a harmless text file.