What Are Claude Skills? How They Work (2026 Guide)
A Claude skill is a folder of instructions loaded only when needed. How progressive disclosure works, where skills run, the four types, and why the open standard matters.
What a Claude skill actually is
A skill is a folder. At minimum it contains one file, SKILL.md, with two parts: a small YAML header (the skill's name and a description of when to use it) and a body of markdown instructions. A skill can also ship helper scripts, templates, or reference files alongside it.
The clever part is how Claude loads it. Anthropic calls this progressive disclosure: Claude always sees the name and description of every installed skill, but the full instructions are only pulled into context when a request matches. Ten installed skills cost almost no context until one is actually relevant.
Where skills work
- Claude.ai - on Free, Pro, Max, Team and Enterprise, as long as code execution is enabled in settings.
- Claude Code - drop a folder in .claude/skills/ and it becomes a slash command.
- The API - in beta, via the code execution tool.
The four kinds of skills
- Anthropic skills: built-in skills for Excel, Word, PowerPoint and PDF work. These are what make Claude genuinely good at producing real spreadsheets and documents.
- Custom skills: ones you write yourself and upload.
- Org skills: provisioned by Team and Enterprise owners so a whole company shares the same procedures.
- Partner skills: a directory of skills from companies like Notion, Figma and Atlassian.
Skills vs projects vs MCP vs custom instructions
People mix these up constantly, so here is the honest version:
- Custom instructions are broad and always on. They shape how Claude behaves everywhere, not how it does one task.
- Projects are static context. Everything in a project is loaded all the time, whether relevant or not.
- MCP gives Claude access to tools and data (your database, your CRM, GitHub). It answers "what can Claude reach?"
- Skills give Claude procedural knowledge. They answer "how should Claude do this task?" A skill can even tell Claude how to use your MCP tools correctly.
A rule of thumb: instructions for behavior, projects for knowledge, MCP for access, skills for procedure.
Skills are an open standard
The format Anthropic uses is published as the Agent Skills open standard at agentskills.io. That matters more than it sounds: a SKILL.md file written for Claude is portable. On Krater, the same file can be uploaded in Settings > Skills and used across 400+ models, so a procedure you write once is not locked to one vendor.
Make a skill in one prompt on Krater
Krater supports the same open skill format, and its agent can write the skill for you. Describe the procedure in one message, approve the card, and the skill is live. We tested it: asking for a Shopify product-description skill took 39 seconds and about 2.5 credits, and the finished skill followed its own rules in a fresh chat. See the full walkthrough with the exact prompt in our guide to creating a skill.
We build Krater, an AI platform that supports the same open skill standard covered here. We think being honest about how skills work is better marketing than pretending they only exist on our platform.
FAQ
Do Claude skills cost extra?
No. Skills are available on every Claude plan including Free. Using one just costs the tokens of that conversation.
What are AI skills in general?
The same idea, vendor-neutral: a packaged procedure an AI agent loads when relevant. Anthropic popularized the SKILL.md format and published it as an open standard, which is why "Claude skills" dominates the search results.
Can Claude skills run code?
Yes, a skill can bundle scripts Claude executes through code execution. That is why you should only install skills from sources you trust - read the SKILL.md before enabling anything.
Do skills work in Claude Code?
Yes. A folder in .claude/skills/your-skill/SKILL.md becomes a /your-skill command. Details in our Claude Code skills guide.