First choose what you are installing
Claude Code already ships with bundled Skills such as /code-review, /debug, and /batch. A standalone Agent Skill is a directory with a SKILL.md entry point. A Claude Code plugin is a packaged unit that can include Skills, agents, hooks, and MCP servers.
Use the installation path that matches the thing you want. A standalone Skill does not need to become a plugin, and a plugin is not just another name for one SKILL.md file.
| What you need | How to recognize it | Where to start |
|---|---|---|
| Bundled Skill | /code-review, /debug, or /batch | Try it in Claude Code; do not install a second copy. |
| Standalone Skill | A folder centered on SKILL.md | Use the Skills CLI or copy the complete folder to a Skill path. |
| Plugin | May bundle Skills, agents, hooks, or MCP servers | Open /plugin and review the whole package before enabling it. |
| claude.ai synced Skill | Enabled for your Claude account | Use account sync for Cowork or cloud sessions; a local personal folder is not enough. |
Install a GitHub Skill with the Skills CLI
Run the first command from the repository where you want the Skill. It targets Claude Code explicitly and uses the CLI's default project scope. The second command adds personal scope for a Skill you deliberately want across projects on this machine.
These are concrete examples from the CLI documentation, not a recommendation to install this particular package. Check the repository, the exact Skill folder, and its permissions before you run an install command.
npx skills add vercel-labs/agent-skills --skill frontend-design --agent claude-codenpx skills add vercel-labs/agent-skills --skill frontend-design --agent claude-code -g- Use project scope for a repository-specific workflow or a Skill the team should review with the project.
- Use -g only for a deliberate cross-project workflow. It targets your user directory, not a machine-wide administrator installation.
- Read the installer prompts. Do not add --yes just to skip a source, scope, or install-method choice you have not reviewed.
Manual install: copy the complete Skill folder
If the source cannot produce a reliable CLI command, open the listed repository and copy the Skill's complete directory. Keep SKILL.md together with any referenced scripts, examples, templates, and supporting files; copying only the entry file can leave broken references.
Claude Code uses different paths for repository-level and personal Skills. On Windows, the personal path expands from your user profile; it does not require an administrator account.
| Scope | Claude Code path | Best for |
|---|---|---|
| Project | .claude/skills/<skill-name>/SKILL.md | One repository; commit the folder when teammates should share it. |
| Personal | ~/.claude/skills/<skill-name>/SKILL.md | Your projects on this machine. On Windows, this is under %USERPROFILE%\.claude\skills\. |
- Keep one canonical copy unless you intentionally need both project and personal scope.
- Confirm the final folder contains SKILL.md and every file the instructions reference.
- For portable team setup, prefer the project path and review the committed files like other project instructions.
Verify the path and the active Claude Code session
The CLI listing checks what the installer placed for Claude Code. In an interactive Claude Code session, run /skills to confirm Claude Code can see it. Inspect the directory too when the exact scope matters; the list and the path answer different questions.
Invoke a Skill with its /name, or describe a matching task and let Claude Code choose it when relevant. If you create the top-level .claude/skills directory after a session has already started, run /reload-skills once. Claude Code otherwise watches existing Skill directories for SKILL.md changes.
npx skills ls -a claude-codenpx skills ls -g -a claude-code- The folder name or SKILL.md name matches the command you expect to use.
- The Skill appears in /skills in the same local session where you installed it.
- Test it on a small, reversible task without secrets or production credentials, then review the result and file changes.
Troubleshoot when a Claude Code Skill does not appear
A successful download, a Skill listed by the installer, and a Skill available in the current session are three different checks. Match the symptom before reinstalling; the issue may be the target Agent, command name, or session type.
| What you see | What to check next |
|---|---|
| The CLI lists it, but Claude Code /skills does not | Confirm the install targeted claude-code and the SKILL.md is under .claude/skills/ (project) or ~/.claude/skills/ (personal). If the top-level .claude/skills directory was created after the session started, run /reload-skills. |
| It appears, but the /command name is different | Check the name field in SKILL.md frontmatter. That field supplies the command name; when it is absent, Claude Code uses the Skill directory name. |
| It works in a terminal, but not in Cowork or a cloud session | That is expected for a local ~/.claude/skills/ install. Enable the Skill for your claude.ai account, or commit a project Skill under .claude/skills/ to the repository. |
Why a local Skill may not appear in Cowork or the cloud
Cowork and cloud sessions do not read ~/.claude/skills/ from your computer. They load Skills enabled for your claude.ai account; cloud sessions can also load project Skills committed under .claude/skills/ in the repository. A personal -g install is not a cross-machine or cloud sync.
If a package includes hooks, MCP servers, or other components, it may be a plugin rather than a standalone Skill. Use the plugin flow for that package and review every component it enables.
- For a local terminal session, verify the local path and the Claude Code /skills list.
- For Cowork or cloud, enable the account-synced Skill or use a committed project Skill, according to your team's policy.
- For a plugin, inspect its Skills, agents, hooks, and MCP servers before enabling it; do not assume it only adds text instructions.
Review the source before trusting the result
A ranking, install count, or successful setup does not establish what a Skill will do in your repository. Read its instructions and referenced files, note the source revision, and check what file, command, network, or credential access its first task would need.
Keep the first run narrow and reversible. Treat its output as evidence to review, not as permission to publish, merge, or change production systems.
- The source repository and Skill directory are the ones you intended to install.
- Referenced scripts and files are understandable and relevant to the task.
- The first test has clear boundaries, a human-reviewable result, and a rollback path.