How to Install Claude Skills: Code, Desktop, Web (2026)

Installing a Claude skill in Claude Code takes three commands:

git clone https://github.com/anthropics/skills
cp -r skills/document-skills/docx ~/.claude/skills/
# restart Claude Code — skills are scanned at startup

That’s the whole install. The folder lands in ~/.claude/skills/, Claude reads the SKILL.md inside it on the next launch, and from then on the skill activates automatically when a matching task comes up.

If it were always that easy, this article would end here. It doesn’t, because the install is the reliable part. What fails is everything around it: wrong directory nesting, forgotten restarts, and above all skills that install perfectly and then never fire. Of the skills we’ve tested for the SkillProof catalog, roughly 45% had some kind of trigger problem out of the box. We install and run every skill before listing it, so this guide is built from those test logs, not from reading READMEs.

If you’re not yet sure what a skill even is (versus an MCP server or a system prompt), read What are Claude skills first and come back. This guide assumes you have one picked out and want it working.

What you need before installing

The list is short.

For Claude Code you need Claude Code 2.x installed and git on your PATH. That’s it. Skills are plain folders with markdown inside; there’s no package manager, no build step, no dependency resolution. The one exception is skills that call external tools. Webapp Testing needs Playwright installed locally, which is why it scored 4/5 on install ease in our tests instead of 5/5. The skill’s page will tell you about requirements like that; the skill itself often won’t until it fails mid-task.

For Claude Desktop and claude.ai you need a paid plan (Pro, Max, Team or Enterprise) and code execution enabled in settings, because skills run inside the code execution sandbox there. You also need the skill as a zip file rather than a folder. More on that below.

One thing you don’t need: trust in the README. Authors describe what they hoped the skill does. Before installing anything from an unfamiliar repo, open its SKILL.md and skim it. It’s the entire skill. If the instructions look thin or the frontmatter description is one vague sentence, you’ve just previewed your trigger problems.

Installing skills in Claude Code

Every skill is a folder containing a SKILL.md file, optionally with reference files and scripts alongside it. Installing means putting that folder in one of two places.

Global vs project: the claude code skills folder

The global folder is ~/.claude/skills/. Skills here are available in every project on your machine. This is the right home for general-purpose skills: document generation, debugging discipline, writing helpers. Our take: 80% of what you install belongs here, because most skills encode a capability, not a project convention.

The project folder is .claude/skills/ inside a repository. Skills here only activate within that project. Use it for team conventions, project-specific config, or anything you want teammates to get automatically, because project skills travel with the repo. Someone clones it, they have the skills. No onboarding doc required.

When the same skill name exists in both, the project copy wins. That’s a feature (a project can override your global defaults) and occasionally a bug you’ll spend twenty minutes on, which we cover in troubleshooting.

Paths on macOS, Windows and Linux

macOS and Linux use the same path: ~/.claude/skills/ for global, <project>/.claude/skills/ for project scope. On macOS that expands to /Users/<you>/.claude/skills/; on Linux, /home/<you>/.claude/skills/.

On Windows the global folder is %USERPROFILE%\.claude\skills\, which typically means C:\Users\<you>\.claude\skills\. In PowerShell, ~ resolves the same way, so ~\.claude\skills\ works there too. Note the folder starts with a dot on every platform. Windows Explorer hides it by default and macOS Finder does too (press Cmd+Shift+. to show hidden files), which is why we recommend doing installs from the terminal even if you live in a GUI.

The final structure must be exactly this:

~/.claude/skills/
  docx/
    SKILL.md
  frontend-design/
    SKILL.md

Skill folder, then SKILL.md directly inside it. Not skills/docx/docx/SKILL.md. That double-nesting is the single most common install mistake we see, and Claude fails silently on it: no error, the skill just doesn’t exist as far as Claude is concerned.

How to install a claude skill from github

Most skills live on GitHub, usually several to a repo. The pattern is: clone the repo somewhere temporary, copy the one folder you want into your skills directory, delete the clone if you like.

Here’s a real example using Anthropic’s official repo, installing DOCX, the Word-document skill that scored 9/10 on output quality in our tests:

git clone https://github.com/anthropics/skills
cp -r skills/document-skills/docx ~/.claude/skills/

Restart Claude Code. Done. The same repo carries Frontend Design, the highest-scoring skill in our design category (10/10 output; the before/after difference on UI work is the largest we’ve measured):

cp -r skills/frontend-design ~/.claude/skills/

Notice the paths differ: document-skills/docx is nested one level deeper than frontend-design. Repos organize skills however the author felt like that day. Always check where the SKILL.md actually sits and copy the folder that directly contains it. Every listing in our documents category and elsewhere in the catalog shows the exact copy command pre-filled, because we got tired of working these paths out by hand ourselves.

No git? Download the repo zip from GitHub (Code → Download ZIP), extract it, and move the skill folder into place. Same result, more clicking.

Plugin marketplaces

Some authors distribute through Claude Code’s plugin system instead of bare folders:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers

That’s the real install command for the Superpowers collection, home of Test-Driven Development, which scored 9/10 on output in our tests and 5/5 on trigger reliability. Plugins handle updates for you, which folder copies don’t, and in our experience plugin distribution correlates with maintenance discipline: an author who set up a marketplace tends to be an author who ships fixes. When a skill offers both install paths, take the plugin.

Installing in Claude Desktop and claude.ai

Claude Desktop and claude.ai don’t read your filesystem, so the folder method doesn’t apply. Instead you upload the skill as a zip through the capabilities settings.

The steps, current as of mid-2026:

1. Zip the skill folder so that SKILL.md sits at the top level inside the archive (zip the folder itself, not its parent).
2. In Claude Desktop or claude.ai, open Settings, then Capabilities.
3. Enable code execution if it isn’t already on. Skills won’t appear as an option without it.
4. Under Skills, choose upload and select your zip.

The skill is then available in new conversations. Anthropic’s curated skills (the document set, mostly) can also be toggled on directly in that same settings area with no upload at all, and if the skill you want is there, use the toggle. It’s versioned and maintained for you.

Two honest caveats from our testing. First, skills that shell out to local tools obviously can’t do so in the claude.ai sandbox; a skill that drives Playwright or reads your git history is a Claude Code skill, full stop. Second, upload-based installs are per-account, not per-machine, which is convenient right up until you’re maintaining different versions in Claude Code and claude.ai and getting different behavior from “the same” skill. If you work in both, keep Claude Code as the source of truth and re-zip when you update.

How to verify the skill actually triggers

This is the step almost everyone skips, and it’s where installs actually fail. A skill can be perfectly placed, perfectly formatted, and never activate once, because activation depends on Claude matching your request against the description field in the skill’s frontmatter. If that description is vague, the skill is furniture.

We test trigger reliability explicitly, and it’s the weakest link in the ecosystem: around 45% of the skills we’ve put through testing had trigger problems, from “fires only if you quote the skill name” to “never fires at all.” It’s why trigger reliability is a separate 5-point score on every SkillProof listing rather than something we fold into a general grade.

Verification takes two minutes:

1. Restart Claude Code (or start a fresh conversation on claude.ai).
2. Ask Claude directly: “What skills do you have available?” A correctly installed skill appears in the list. If it doesn’t, you have a path problem, not a trigger problem; go back to the folder structure.
3. Give it a task that should obviously trigger the skill, phrased naturally, without naming the skill. For DOCX: “Create a Word document summarizing this project.” Watch whether Claude announces it’s using the skill.
4. Then phrase the same task the way you’d actually type it on a Tuesday afternoon. This is the test that matters. Plenty of skills pass step 3 and fail step 4.

If the skill shows up in step 2 but won’t fire in step 4, the fix is usually editing the description in its frontmatter to include the words you actually use. You’re allowed to do that; it’s your copy. Our free skill validator checks a SKILL.md for the frontmatter mistakes that cause weak triggering, and it will flag a description that’s too short or too generic before you waste a session finding out the hard way.

For skills with a trigger score below 5/5 in our catalog, the test notes on the listing say exactly what phrasing made them fire. That’s often faster than fixing the frontmatter yourself.

Updating and removing skills

There’s no update command for folder-installed skills. Updating means replacing the folder:

cd skills && git pull
cp -r document-skills/docx ~/.claude/skills/

If you deleted the clone after installing, clone again and copy again. This is genuinely a weak point of the folder approach; nothing tells you an update exists. Plugin-installed skills do better, since the plugin system manages versions. It’s one more reason to prefer the plugin path when offered.

A practical note if you’ve customized a skill’s frontmatter to fix its triggering: a blind copy will overwrite your edits. Diff before you replace. We keep customized skills in a small git repo of our own and merge upstream changes deliberately, which sounds like overkill until the first time an update silently reverts a fix.

Removing a skill is the one part of this whole system with zero gotchas:

rm -rf ~/.claude/skills/docx

Restart Claude, gone. Skills don’t touch system state, don’t register anything, don’t leave config behind. On claude.ai, remove the skill from the capabilities settings. That’s it.

Troubleshooting

Four failure modes cover nearly every “I installed it and nothing happened” report we’ve received.

The skill never fires. First check it’s listed when you ask Claude what skills it has. If it’s listed but dormant, the frontmatter description is too vague for Claude to match your requests against, which is an authoring bug you inherit at install time. Run the SKILL.md through the skill validator, then rewrite the description with concrete trigger phrases: name the file formats, the verbs, the situations. “Helps with documents” fires on nothing. “Use when the user wants to create, edit or convert .docx Word files” fires reliably. If you’re going to be editing frontmatter anyway, our guide to writing your own Claude skill explains what makes a description fire, and the same rules apply to repairing someone else’s.

Name collisions. Two skills with the same folder name, one global and one project-scoped, and the project one wins silently. Or two skills covering the same domain (two code-review skills, say) and triggering becomes a coin flip. Check both ~/.claude/skills/ and the current project’s .claude/skills/ when behavior seems inconsistent. For same-domain overlap, keep one; there’s no priority system to configure, and in our testing, overlapping skills degrade each other even when both are individually good.

It needs a restart. Skills are scanned at startup, not watched. Every install, update or removal requires restarting Claude Code before it takes effect. Obvious once you know, invisible until you do. We’d estimate a third of the “broken install” reports we get are just this.

Permissions. If Claude lists the skill but errors when using it, check that the files are readable (chmod -R u+r ~/.claude/skills/<name> on macOS/Linux) and that any bundled scripts are executable. Folders copied from zip downloads or synced through Dropbox-type tools occasionally arrive with mangled permissions. On Windows, files downloaded from the internet can carry a blocked flag; right-click, Properties, Unblock.

And a fifth, half a failure mode: skills with real prerequisites. Brand Guidelines installs cleanly but does nothing useful until you edit its config with your palette and fonts, which is why it scored 3/5 on install ease despite being a good skill. When a listing on our site says “works with setup,” this is what it means. Read the test notes before assuming the skill is broken.

Installing 10 skills at once

The copy-paste method scales linearly and painfully: ten skills means ten clones or ten copies, ten trigger checks, and a real chance two of them overlap. There’s no official bulk installer.

Our answer is SkillProof bundles: curated sets of tested skills that install with a single command. Each bundle ships a verified config where every skill passed our testing individually and the set was checked for trigger conflicts as a group, since “ten good skills” and “ten skills that work together” are different claims. If you’d rather assemble your own stack, the free path is a personal git repo containing your chosen skill folders and a one-line cp -r my-skills/* ~/.claude/skills/. You lose the conflict testing, but you gain a reproducible setup you can put on a second machine in thirty seconds.

Either way, resist installing twenty skills speculatively. Skill descriptions all get read when Claude decides what to activate, and a crowded skills folder full of vague descriptions makes triggering worse for everything. Install what you’ll use this month. Add more when a task demands it.

FAQ

Where do Claude skills get installed? In Claude Code: ~/.claude/skills/ for global skills, <project>/.claude/skills/ for project-scoped ones. On Windows, %USERPROFILE%\.claude\skills\. On claude.ai and Claude Desktop, skills are uploaded as zips under Settings → Capabilities and stored with your account, not on disk.

Do I need to restart Claude Code after installing a skill? Yes, every time. Skills are scanned once at startup. Installs, updates and removals all take effect on the next launch, and this is the most common reason a fresh install “doesn’t work.”

Can I install skills without git? Yes. Download the repo as a zip from GitHub, extract it, and copy the skill folder into ~/.claude/skills/. Git only makes updating easier; it’s not required for anything.

Why doesn’t my installed skill ever activate? Assuming the path is right, the frontmatter description in its SKILL.md is too vague for Claude to match your requests. Roughly 45% of skills we test have this problem. Edit the description to include concrete trigger phrases, or check our skill validator and the skill’s SkillProof test notes for known workarounds.

Are Claude skills safe to install? A skill is instructions plus optional scripts that Claude can execute, so treat it like any code you run: read the SKILL.md before installing, be wary of unmaintained repos, and prefer skills that have been tested by someone. Every skill in our catalog was installed and run by us before listing, and the verdict on the listing reflects what actually happened.