Claude Skills: what they are and how to use them

A Claude Skill is a folder containing a set of instructions, and sometimes scripts or templates, that Claude reads on its own when a request matches what the Skill is for. You don't run it like a slash command. You ask for the thing you want in plain English, Claude recognizes the match, and it follows the packaged instructions instead of you re-explaining them.

How Claude actually decides to use one

Every Skill you have installed is built around one required file: SKILL.md, a short Markdown file with a YAML header (name and description) followed by instructions in plain English. Anthropic's own docs describe Skills as "instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks," and the loading part is the interesting bit.

“Claude keeps every installed Skill's name and description in view at all times, but only pays the full cost of one when a task actually calls for it.”

That's done in three tiers. First, just the name and description of every Skill sit in the background at all times, a tiny fraction of a token budget each. Second, the moment your request matches a Skill's description, Claude reads the full body of that SKILL.md file, still a small amount of text.

Third, if that file points to extra material, a reference doc, a script, a template, Claude only opens those on demand, and code inside a script never enters the conversation itself, only its output does. This is why installing ten Skills doesn't slow every single chat down: most of a Skill's weight sits dormant until it's actually relevant.

Claude runs all of this inside a sandboxed environment with filesystem access and the ability to execute code, navigating a Skill's folder the way you'd read an onboarding binder: skim the summary, open the relevant page, run the attached tool if there is one.

Skills shipped on October 16, 2025, as "Agent Skills," launching with four pre-built ones for PowerPoint, Excel, Word, and PDF, available across claude.ai, Claude Code, and the API. On December 18, 2025, Anthropic published the format as an open specification at agentskills.io, and within days tools outside Claude, including VS Code, OpenAI's Codex CLI, and roughly thirty others, added support for the same SKILL.md format.

Skills vs. Projects vs. Artifacts

This is the mix-up almost everyone runs into first, so it's worth being precise about it using Anthropic's own definitions:

• Skills are specialized procedures that "activate dynamically when needed." Conditional, task-specific, and only loaded when relevant.
• Projects are static background knowledge: files and instructions you attach once, then loaded into every conversation inside that Project whether or not it's relevant to the message you just sent.
• Custom instructions sit above both: general preferences applied broadly across all your conversations, not tied to a specific task.

There's a second mix-up worth flagging even though it's less common for a total beginner: Skills vs. MCP. MCP is infrastructure, a live connection between Claude and an outside system, like your calendar or a database. A Skill is knowledge, instructions for how to use a connection correctly once it exists, or how to do a self-contained task with what Claude already has locally.

A Skill can bundle a script for a deterministic local job, but it can't reach a live API or database at runtime the way an MCP connection can. That's a real capability boundary, not just different names for the same idea.

How you'd actually run into one

Support differs by surface, and it's not symmetrical:

On claude.ai specifically, here's the actual first-run path:

Team and Enterprise admins get one more layer: an org owner can turn Skills on org-wide and centrally provision specific ones as defaults, though individual users can still switch any admin-provisioned Skill off for themselves.

Real examples worth knowing about

PowerPoint, Excel, Word, and PDF are the four pre-built Skills that shipped with the feature. Ask for a presentation, a spreadsheet with formulas, a formatted document, or a filled-out PDF form, and Claude reaches for the matching one without you naming it.

Partner Skills in the Skills Directory (launched December 18, 2025, browsable at claude.com/connectors) are commercially built and org-installable: Canva for generating on-brand presentations and campaigns from one prompt, plus Notion, Figma, Atlassian, Cloudflare, Sentry, Vercel, Zapier, and Ramp.

The Claude API Skill is open-source and bundled with Claude Code, giving Claude current API reference material and SDK docs for eight programming languages. It's also installable on its own from Anthropic's public skills repo.

Community-built Skills show up on the same public repo and third-party roundups: a "Frontend Design" Skill that pushes Claude to commit to one bold visual direction before writing code, a "Superpowers" Skill that structures a coding agent's brainstorm-to-review workflow, and a single-file behavioral-guidelines Skill from AI researcher Andrej Karpathy aimed at common agent failure modes. Treat these as examples of what people build, not as a ranked or verified popularity list.

A workflow you keep re-explaining. The actual pitch for a working professional: a weekly status update in a set format, customer feedback triage, sales-call prep notes, brand-compliant document formatting, standardized technical docs. Write the instructions once as a Skill, and it fires on matching requests going forward instead of you retyping them.

Who this is really for

Custom Skills on claude.ai genuinely don't require code. A Skill "can be as simple as a few lines of instructions," pure Markdown, no script required: write a SKILL.md with a name and description in the YAML header, plus instructions in the body, zip the folder, upload it under Customize → Skills, and test it. If Claude isn't firing the Skill when you expect, the description is almost always the fix, since that's literally the text Claude reads to decide relevance.

The name field is capped at 64 characters (lowercase letters, numbers, hyphens only), the description at 1024, and Anthropic recommends keeping the body itself under 500 lines.

Claude Code and API Skills are a different story: filesystem paths, YAML frontmatter you're expected to get exactly right, and for the API, two beta headers plus the code-execution tool. That side is developer territory.

A few honest limits worth knowing before you build around this:

• Custom Skills don't sync across surfaces. One uploaded to claude.ai isn't available via the API, and Claude Code Skills are separately filesystem-based from both. You manage each surface on its own.
• Sharing is narrower than you'd guess. On claude.ai, a custom Skill is tied to the individual who uploaded it, even on Team and Enterprise there's no admin-managed way to push a custom Skill to the whole org, only pre-built and Directory partner Skills get that treatment. Every teammate uploads their own copy.
• Network access is inconsistent by design. On the API, a Skill has zero network access and can't install packages at runtime. On claude.ai, a Skill can install from npm, PyPI, or GitHub, and network access itself varies by admin setting. On Claude Code, a Skill gets full local network access, same as any program on your machine. The same Skill can behave differently depending on where you run it.
• Skills aren't covered by Zero Data Retention. If you're on an API plan relying on ZDR elsewhere, Skill definitions and execution data still follow Anthropic's standard retention policy.

Tested on Claude Sonnet 5, July 2026: asking for a slide deck without mentioning "Skill" or "PowerPoint Skill" by name was enough for Claude to invoke the pptx Skill and produce a downloadable file, matching the "just ask normally" behavior described in Anthropic's own help docs.

This site isn't run by Anthropic. We're an independent guide to using Claude well, and when a feature has a real gap, like custom Skills not carrying over between claude.ai and the API, we say so instead of smoothing it over.

What are Claude Skills?

Folders of instructions, and optionally scripts or templates, that Claude loads automatically when a request matches what the Skill is built for. They teach Claude to do a specialized task the same way every time, like formatting a document in your brand style, without you re-explaining it.

How are Skills different from Projects?

A Project is always-on background knowledge: files and instructions loaded into every chat in that workspace, relevant or not. A Skill only loads when a specific request matches it, and it packages a repeatable process rather than reference material.

Can I make my own Claude Skill?

Yes, and it doesn't require code. Write a SKILL.md file with a name, a description, and plain-English instructions, zip the folder, and upload it under Customize → Skills on claude.ai. If Claude doesn't trigger it as expected, rewrite the description first.

Do Claude Skills work in Claude Code?

Yes, but differently. Claude Code has no pre-built Skills; custom ones live as folders in ~/.claude/skills/ or a project's .claude/skills/ directory and are discovered automatically. They're separate from anything uploaded to claude.ai or the API.

If you're earlier in the process, how to use Claude covers your first session end to end, and Claude Artifacts explains the panel where a Skill's output usually ends up. For the wider set of write-ups, browse more reviews.