Skills: teaching your tool your way of working

You have explained the same thing to your AI tool four times this month. How your team writes commit messages. The shape of a component in this codebase. The checklist you go through before a release. A skill is where that explanation lives so you stop typing it.

Not the same thing as an MCP server

This is the distinction worth getting straight, because installing the wrong kind of thing is the usual first mistake. An MCP server gives your tool a new ability. It could not read your database; now it can. A skill gives your tool new instructions. It could always write a commit message; now it writes one the way your team does. If the thing you want is "reach something it cannot reach", you want a server. If it is "do the thing it already does, but properly", you want a skill.

Why your tool ignores it

The first line of a skill is a description of when to use it, and that line is doing almost all the work. Your tool reads the descriptions of every installed skill and picks one when the description matches what you asked for. It does not read the skills themselves until it has picked one. So a skill described as "helps with code" never fires, because nothing is specifically that. A skill described as "use when writing or reviewing a database migration" fires exactly when it should.

Install one

Skills come from GitHub repositories, and one repo often holds several. Here is a well-made set to start from, installed one skill at a time.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a claude-code

The skill's files written into .claude/skills. Run it with / and the skill's name.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a codex

The skill's files written into .agents/skills. Run it with / and the skill's name.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a cursor

The skill's files written into .cursor/skills. Run it with / and the skill's name.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a github-copilot

The skill's files written into .github/skills. Run it with / and the skill's name.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a antigravity

The skill's files written into .agents/skills. Run it with / and the skill's name.

npx skills add https://github.com/anthropics/skills --skill frontend-design -a gemini-cli

The skill's files written into .gemini/skills. Run it with / and the skill's name.

Note the two kinds of tab above. Anything that runs in a terminal reads skills from a folder, so the skills CLI can put them there for you. Claude.ai and ChatGPT take a ZIP through their own settings screen instead — no command exists that would change that, and a skill that depends on running scripts may not work in them at all.

Then write your own

Installed skills are the smaller half. The ones worth having are about your project, and nobody else can write those. The test for whether something should be a skill: have you explained it more than twice, and is it only relevant sometimes? Always-relevant things belong in your instructions file, which loads every session. Sometimes-relevant things belong in a skill, which does not.

The third prompt is the useful one once you have a few. A skill that never fires is not a skill, it is a file.

• Browse skills
• Global or project: where config lives
• MCP servers: giving your AI tool hands
• AGENTS.md: a README for your coding agent