Best Claude Skills and How to Install Them

Anthropic publishes around twenty Skills in a public repository, and the useful ones split cleanly by what they change. Some fix output formats Claude is otherwise mediocre at — real .docx files, real slide decks, real spreadsheets. Some install a process Claude follows instead of improvising. One builds more Skills.

If you install nothing else, install the document Skills. They're the only ones that change what Claude can physically produce rather than how it behaves.

Here's the short list, then the three ways to install them.

Quick comparison

One licensing note worth knowing before you build on these: most of the repository is Apache 2.0, but the document Skills — docx, pdf, pptx, and xlsx — are source-available rather than open source. If you're planning to redistribute a derivative, read the terms rather than assuming.

Anthropic also labels the repository as demonstration and educational material, and warns that what ships in Claude may differ from these implementations. Treat them as strong reference patterns, not as a supported product surface.

The document Skills are the ones that matter most

Ask Claude for a Word document without a Skill and you get markdown, or a file that opens with broken formatting. The docx, pptx, xlsx, and pdf Skills bundle the actual libraries and conventions for generating those formats, so what comes back opens cleanly in Office.

This is a bigger deal than it sounds for anyone whose work ends in a deliverable someone else has to open. A report that needs reformatting isn't finished; a report that opens correctly is. The gap between those two is the entire value of the Skill.

xlsx is the strongest of the four, because spreadsheets have the most structure to get wrong — formulas, number formats, multiple sheets, column widths. It's the difference between data you can hand to someone and data you have to rebuild.

Install all four together:

/plugin install document-skills@anthropic-agent-skills

skill-creator: the one that compounds

skill-creator writes Skills. That sounds circular until you've tried writing one by hand and discovered that the hard part isn't the format — it's knowing what belongs in the main file versus a reference file, and how to write a description that makes Claude load the Skill at the right moment and ignore it otherwise.

That description field is the whole game. Skills use progressive disclosure: Claude reads the name and description of every installed Skill cheaply, and only loads the full instructions when one looks relevant. A vague description means the Skill either never fires or fires constantly. skill-creator gets this right by default, which saves several rounds of trial and error.

If you're going to write Skills for your own repeated work — and that's where the real return is — this is the first one to install. Our explainer on how Claude Skills work covers the progressive disclosure mechanism in more depth.

mcp-builder and webapp-testing: the technical pair

mcp-builder scaffolds MCP servers — the connectors that let Claude reach a database, an internal API, or a tool that has no integration yet. Writing one from scratch means learning the protocol's shape first; the Skill encodes a working pattern so you're editing rather than starting.

webapp-testing drives a browser against a running application. Claude opens pages, clicks things, fills forms, and reports what broke. For anyone maintaining a web app, this converts "write tests eventually" into something that happens during a normal session.

Both are in the example-skills plugin:

/plugin install example-skills@anthropic-agent-skills

How to install Claude Skills

Three routes, depending on where you use Claude.

Claude Code: the plugin marketplace

The shortest path. Add Anthropic's repository as a marketplace, then install the sets you want:

/plugin marketplace add anthropics/skills

Then either browse and install interactively, or install a set directly with the /plugin install commands above. Anthropic's Claude Code Skills documentation covers the full command surface, and our guide to Claude Code plugins covers how marketplaces work more generally.

Claude Code: a folder on disk

Skills are directories. Claude Code reads them from two locations — ~/.claude/skills/ for Skills available in every project, and .claude/skills/ inside a project for Skills specific to it.

So installing a Skill you found elsewhere is just putting the folder in the right place and starting a new session. Project-level Skills committed to a repository travel with it, which is how a team shares a review checklist or a deploy procedure without anyone configuring anything.

Claude.ai and the API

On Claude.ai, Skills upload through the web interface rather than a file path. Through the API, Skills attach to requests programmatically — the Agent Skills documentation covers that route.

The underlying format is identical across all three. A Skill is a folder with a SKILL.md file containing YAML frontmatter — a name and a description — followed by instructions in plain markdown. Supporting files sit alongside it and get read only when needed.

That's the whole specification. The reason Skills spread quickly is that there's almost nothing to learn.

What to install first, by what you do

You produce documents other people open. The four document Skills, and nothing else until you've used them.

You have a process you re-explain constantly. skill-creator, then write that process as a Skill once.

You build things. mcp-builder and webapp-testing.

You care how output looks. brand-guidelines and theme-factory for consistency, frontend-design for UI work.

The failure mode is installing everything. Each Skill's description competes for Claude's attention, and a pile of Skills you don't use makes the ones you do use fire less reliably. Install four, use them for a week, add more when a gap shows up.

Where Skills stop

A Skill teaches Claude how to do something. It doesn't give Claude anything new to do it with — no database access, no internal API, no files outside what the session can already reach. That's MCP's job, and the two compose rather than compete. Anthropic's engineering write-up on Agent Skills is the clearest statement of where the boundary sits.

A Skill also assumes you're already running Claude Code or Claude with somewhere to put the folder. That's the barrier for most people — not the Skill, the environment around it.

If that's where you stall, Taku mirrors a working AI setup into a desktop workspace and runs it, so a configuration someone else proved out becomes something you can use without reproducing their machine first. The free app library is the fastest way to see what that looks like. Taku is in Beta, and the Mac app is available now.

FAQ

How do I install Claude Skills?

Three ways. In Claude Code, add a plugin marketplace and install from it, or drop the Skill folder into ~/.claude/skills/ or a project's .claude/skills/. On Claude.ai, upload through the web interface. Through the API, attach Skills to requests.

Where does Claude Code look for Skills?

~/.claude/skills/ for Skills available everywhere, and .claude/skills/ in a project root for Skills scoped to that project. Both are picked up when a session starts.

What's inside a Skill?

A folder with a SKILL.md file. That file has YAML frontmatter with a name and description, then markdown instructions. Supporting files can sit alongside and are loaded only when relevant.

Are Anthropic's official Skills open source?

Mostly. The repository is largely Apache 2.0, but the docx, pdf, pptx, and xlsx document Skills are source-available rather than open source. Check the license before redistributing.

Can I install too many Skills?

Yes, in a practical sense. Every installed Skill's description is in play when Claude decides what to load, so a large pile makes selection noisier. Keep the set close to what you actually use.

Do Skills work across Claude Code, Claude.ai, and the API?

The format is the same in all three. How you install differs — plugin or folder in Claude Code, upload on Claude.ai, programmatic attachment through the API.

Do I need a Skill to use MCP?

No. They're independent. Skills carry instructions; MCP carries connections. Using both together is common, but neither requires the other.