Install in 60 seconds
How to install a Claude Code skill
Copy a markdown file into ~/.claude/, restart Claude Code, type /<name>. That's the whole thing.
Guide verified April 21, 2026
The short answer
Skills are plain markdown files. Save them to ~/.claude/commands/ (slash commands), ~/.claude/agents/ (subagents), or ~/.claude/skills/<name>/SKILL.md (bundled skills with scripts). Restart Claude Code. Invoke with /<name>.
Directory cheat sheet
Which directory does what — skim this once, bookmark if needed.
| Directory | Holds | Invoke with |
|---|---|---|
| ~/.claude/commands/ | Slash commands — named prompts you run manually. | /<name> |
| ~/.claude/agents/ | Subagents — specialists Claude can delegate to. | /<name> or auto |
| ~/.claude/skills/<name>/ | Bundled skills — SKILL.md plus supporting scripts or docs. | auto via frontmatter |
| <project>/.claude/ | Per-project overrides — same structure, project-scoped. | same |
Step-by-step
From catalogue click to working skill in six steps.
- 1
Pick a skill from the catalogue
Browse /claude-code-skills and open the one you want. The High-signal tier is where to start — those skills earn their keep daily across many projects.
- 2
Copy the skill's content
On the skill's page, the setup block shows the exact filename and path. For most entries, you'll copy a markdown file from the source repo (link on the page) into your own .claude/ directory.
- 3
Save to the right directory
The path depends on what the skill is:
# Slash command (most skills) ~/.claude/commands/<name>.md # Subagent (parallel specialist) ~/.claude/agents/<name>.md # Bundled skill (with scripts / extra files) ~/.claude/skills/<name>/SKILL.md # Per-project equivalents (override globals) <project-root>/.claude/commands/<name>.md <project-root>/.claude/agents/<name>.md <project-root>/.claude/skills/<name>/SKILL.md
- 4
Preserve attribution
If the skill came from a community repo (not your own work), keep a one-line comment at the top crediting the source. The catalogue's licence gate (MIT / Apache-2.0 / BSD / CC0 / Unlicense) permits reuse, but MIT and Apache require you keep the original copyright notice.
<!-- Source: <author> / <repo> — MIT --> <!-- https://github.com/<author>/<repo> -->
- 5
Restart Claude Code
Quit the current session (Ctrl+C twice) and run claude again. New skills are registered at session start, not mid-session.
claude
- 6
Invoke it
Type / in Claude Code — the skill appears in autocomplete. Some skills auto-trigger based on context (check the frontmatter's description field). Most fire when you type /<name> and press enter.
You: /commit Claude: [runs git status / diff / log, drafts message, commits, pushes]
A realistic first five
Start here. These are the most-used High-signal skills from the catalogue — the ones most people keep forever once they install them.
Commit Helper
Stages relevant changes, writes a conventional-commit message, and pushes — in one command.
PR Open
Opens a pull request with a structured description drawn from the full commit range, not just the latest commit.
PR Review
Runs a multi-angle review on an open PR: correctness, style, tests, security, and docs.
Security Audit
Sweeps a repo for OWASP-top-10 issues, hardcoded secrets, insecure dependencies, and auth logic gaps.
Site Health Check
Runs a battery of drift/regression checks on a site — CTA hygiene, data freshness, link health, cruft in pages/.
19 High-signal skills total · browse all
Common mistakes
Saving the file with a .txt extension
Claude Code only reads .md files from these directories. If your editor adds .txt, rename before restart.
Forgetting to restart
Skills are loaded once at session start. A new file dropped mid-session won't appear until the next claude invocation.
Skipping attribution
MIT and Apache-2.0 both require you keep the copyright notice when redistributing. For your own ~/.claude/ that's a nicety; if you commit a skill into a project repo, it's a licence requirement.
Running a bundled skill without reading the scripts
Slash commands are just prompts — safe. Bundled skills can contain shell scripts that run with your permissions. Open the script before saving.
Frequently asked
Where do skills live on my machine?+
Slash commands go in ~/.claude/commands/<name>.md — one file per command. Subagents go in ~/.claude/agents/<name>.md. Bundled skills (with scripts or additional files) go in ~/.claude/skills/<name>/SKILL.md. All three are picked up automatically by Claude Code when it starts. Per-project versions live at .claude/commands/ (etc.) inside the repo and override the global ones for that project.
Do I need to restart Claude Code after adding a skill?+
Yes. Claude Code reads the commands, agents, and skills directories at session start. Quit (Ctrl+C twice) and run claude again. New skills appear when you type / — the autocomplete will include them. If a skill doesn't show up, check the file is .md, the frontmatter is valid YAML, and you saved to the right directory.
How do I invoke a skill once it's installed?+
Slash commands and subagents both run by typing /<name> in Claude Code. Some skills auto-trigger based on context (set via the allowed-tools or description field in frontmatter) — those fire without being called directly. You can check a skill's trigger mode in its frontmatter. When in doubt, type / and browse the list.
How do global skills differ from per-project skills?+
Global skills in ~/.claude/ apply everywhere you run Claude Code. Per-project skills in ./.claude/ (relative to the project root) only apply inside that project, and override global skills of the same name. Use global for personal defaults (commit style, review habits) and per-project for stack-specific workflows (Next.js lint, Prisma migrations, repo-specific CI rules).
What if the skill needs a dependency like gh CLI?+
Install the dependency first. Most of the skills in this catalogue require at least git and gh (GitHub CLI). Some need jq, rg (ripgrep), or specific MCP servers. The setup block on each skill page lists requirements. Claude Code will surface the error if a dependency is missing, but it's faster to pre-install rather than debug at runtime.
Is it safe to install skills from GitHub strangers?+
Read the skill file first. A slash command is just a markdown prompt — low risk. A bundled skill may contain shell scripts that run with your permissions; read those carefully. This catalogue only includes skills from repos with permissive licences, but 'permissive licence' isn't the same as 'audited code'. Trust, but verify: open the file, read what it does, then save it.
Can I edit a skill after I install it?+
Yes — skills are just markdown files. Edit in place to match your stack, conventions, or tone. Fork the file with a new slug if you want to preserve the original alongside your edit. Respect the licence: MIT and Apache-2.0 both require you keep the original author's copyright notice when you redistribute, but local edits in your own .claude/ are fine.
Ready to write your own?
Once you've installed a few and spotted a repeat pattern in your own work, it's time to build one. The template walkthrough covers the frontmatter, body structure, and testing loop.