What Are Claude Skills?

Short definition

A Claude Skill is a folder containing a SKILL.md file: YAML frontmatter with a name and a description that says when to use it, followed by the instructions Claude should follow. Optional subfolders hold scripts, references, templates, and assets that Claude reads only when the task calls for them.

Claude scans the descriptions of available skills while it works, loads the matching one, and then pulls in deeper files as needed. That is the whole mechanism. A skill is not just a prompt snippet, because it can carry files and executable helpers, and it is not a remote service, because the unit is a package you can open and read.

How to find, install, or write one

The fastest way to write a first skill is to ask Claude to do it with the official skill-creator skill, which interviews you about the workflow and generates the folder. Review what it produces before sharing it.

Why Skills exist

General instructions are brittle when the task has many steps, special file formats, house style rules, or repeated quality checks. Skills let a team put that procedure in one named package, so a future agent run can load the same guidance without forcing the user to paste a long prompt every time.

This is especially valuable for document conversion, brand-guideline work, visual design workflows, internal communications, API usage patterns, and other tasks where "do the job our way" matters as much as the raw model capability.

Where they run

Claude Code supports skills from three broad locations: project skills in .claude/skills, personal skills in ~/.claude/skills, and plugin-distributed skills from installed marketplaces. Claude apps expose Anthropic-managed skills and custom skills through settings for eligible plans. The Claude API exposes skills as uploadable, versioned packages for Messages API workflows.

That means the same concept spans several product surfaces, but you should still read the current docs for the surface you are using. Claude Code path rules are not the same thing as Claude app upload rules, and API packaging has its own lifecycle.

Which model runs a skill

The skill format is model-agnostic: the same folder works across the current lineup. As of September 2026 that lineup is Claude Opus 5 as the general default, Claude Fable 5.1 for demanding reasoning and long-horizon agentic runs, Claude Sonnet 5 for the speed and intelligence balance, and Claude Haiku 4.5 as the fastest and cheapest option.

Model choice still changes behavior in two ways worth testing. Stronger models follow long procedures and multi-file routing more faithfully, so a skill that needs Opus 5 or Fable 5.1 to work should say so in its instructions. Reasoning effort also differs by surface — Fable 5.1 defaults to high effort in Claude Code and medium in the Claude apps — which can change how much of a bundled reference the model actually reads. For model selection and pricing, see Claude Central; this guide stays on packaging.

What a Skill is not

A skill is not an MCP server. MCP connects Claude to external tools and data sources through a protocol. A skill can explain how to use a tool, or include a script, but it does not turn into a live connector just because it exists.

A skill is not a complete product integration by itself. It can encode a workflow, provide templates, and run bundled helpers, but credentials, permissions, network access, and external systems still need ordinary engineering review.

FAQ