Official Guide: How to Use Skills in Claude Code

Published on

Anthropic released a “definitive guide” on the trendy technology, “Skills”, very common for those who use Claude Code and agentic programming tools.

Basically, a “skill” is a set of instructions, like well-done Prompt Engineering, that teaches Claude Code how to handle very specific tasks or workflows.

Guide on how to create skills

How it works behind the scenes is actually pretty simple. The skill folder contains a mandatory SKILL.md file, which has the instructions in Markdown format and a YAML header saying what it does. You can also add optional folders with scripts (Python, Bash), reference documents, etc.

From there, it works almost like function calling, with the model having access to all the skills and choosing the one it thinks (or the one you point to) makes sense for some specific use.

The architecture has a really cool concept called “Progressive Disclosure”, with three levels. The first level is the YAML header, which always stays in the system prompt so Claude knows when it should activate that skill. The second level is the body of the Markdown file, which is loaded (disclosed.. get it? get it?) only when Claude thinks the skill is actually relevant to the task. And the third level is the linked files, which it only discovers and navigates if strictly necessary. This helps minimize the context window’s token usage.

And of course, since it comes from Anthropic, it comes integrated with MCP (Model Context Protocol) out of the box. In the guide, they even use a really nice analogy: MCP provides the “professional kitchen”, giving access to tools, ingredients and equipment. Skills, on the other hand, provide the “recipes”, that is, the step-by-step instructions on how to make something.

You can have as many skills as you want in your project. But as I showed in a post last week, the folks behind SkillsBench showed that, at least nowadays, the best thing is to have 2-3 extremely specific and well-made skills, curated by humans (skills created by LLMs didn’t show a big gain).

And to close with a flourish, I also found an awesome repository about exactly this on GitHub, “Everything Claude Code”. The guy (an Anthropic Hackathon winner) built a complete performance optimization system for AI agents, and it already offers loooots of ready-made skills, commands, optimized memory and a focus on security.

And I talked a lot about Claude and Claude Code, but skills also work in Cursor and other agentic programming tools. It’s a huge help and it’s definitely worth studying the code to get inspired and, who knows, start creating your own skills.