Skip to main content
Most of what you know about your job has never been written down. The way your team formats a memo. The rule about always excluding test accounts from revenue queries. The voice your customer emails are supposed to have. Every time you work with Claude, you either re-explain all of that or you accept output that gets it wrong. A skill fixes this. It is a folder of instructions, examples, and optional scripts that Claude discovers and reads on its own, exactly when the task calls for it. You write down how the work should be done once. Claude applies it every time after that, in every conversation, without being reminded. Anthropic’s own one-line pitch for the feature is hard to improve on:
claude.com/skills
The Claude Skills product page: Teach Claude your way of working. Turn your expertise, procedures, and best practices into reusable capabilities so Claude can apply them automatically, every time.

The real pitch, from the real page. Expertise in, automatic application out.

​
Why this is the feature worth mastering

Skills are the most human-replicable thing in Claude. Building an agent takes engineering. Wiring a connector takes an admin. A skill takes a text file. If you can write an onboarding guide for a new hire, you can write a skill, because that is literally what a skill is: an onboarding guide that Claude reads at the moment it becomes relevant. They also solve a problem that grows with you. The better you get at describing work to Claude, the more painful it becomes to re-describe it every day. Skills are where good descriptions go to become permanent.
In a prompt

You type the memo format again. It works today and is gone tomorrow.

Lives for: one conversation
In project instructions

The format lives in one workspace. Chats inside it start briefed. Chats outside it do not.

Lives for: one project
In a skill

The format loads itself whenever any conversation turns to memos, and stays out of the way otherwise.

Lives for: everything you do
The third row has a property the other two lack: it is conditional. A skill costs Claude almost nothing until the moment it is needed. That single design decision is why you can install dozens of them, and it gets its own lesson.

​
What this section covers

What a skill actually is

A folder, a file, and the loading trick that makes unlimited expertise affordable.

Using existing skills well

Where skills live, how Claude decides to fire one, and what to do when it will not.

Caveman: a compression skill

A real community skill that cuts token spend by around 75 percent, dissected as a build example.

Building your own

Evaluation-first authoring, and why the description field is the whole ballgame.

Skills vs sub-agents

Package the knowledge or delegate the work. Same itch, different tools.

Lab: Ship one skill

Draft, install, test, and share a working skill before the session ends.

Quiz: Skills, connectors & Cowork

Shared check after Skills and Connectors. Ten instructor-led questions.

Skills were introduced by Anthropic in October 2025 and later published as an open standard. Primary sources: the engineering deep-dive, the Agent Skills documentation, and the anthropics/skills repository.