How to Install Claude Code Skills

Before installing a Skill

Start with the original publisher's repository or package. Confirm that the URL points to the intended skill, inspect its license and recent changes, and read every file capable of running code. A public repository is not automatically safe. Pay particular attention to shell commands, dependency installation, network access, tool allowlists, and instructions that modify or delete files.

The folder you install should contain SKILL.md at its root. Supporting files must retain their relative paths. Copying only SKILL.md from a multi-file skill can leave broken references or remove safeguards implemented by scripts and templates.

Step 1: Choose personal or project scope

Use personal scope for a workflow you want in all local projects. Use project scope when the instructions describe one repository's commands, architecture, review rules, or deployment process and should be shared through version control.

Do not install repository-specific commands globally simply for convenience. A release skill written for one service can be dangerous or misleading in another.

Step 2: Create the destination directory

For a personal skill on macOS or Linux:

mkdir -p ~/.claude/skills/review-api-change

For a project skill, run this from the repository root:

mkdir -p .claude/skills/review-api-change

In PowerShell, $HOME\.claude\skills\review-api-change is the corresponding personal path. Creating the parent directory is only setup; the skill is not complete until its files are placed inside it.

Step 3: Copy the complete Skill package

The expected minimum is:

review-api-change/
└── SKILL.md

A package may also contain references/, scripts/, assets/, or other files referenced by SKILL.md. Preserve their names and hierarchy. If you clone a repository containing many skills, copy only the specific reviewed skill folder unless the publisher's installation instructions explicitly require the larger package.

For portable Agent Skills, the frontmatter name should use lowercase letters, numbers, and hyphens and match the directory name. Claude Code can derive a command name from the folder in some cases, but following the standard avoids cross-tool ambiguity.

Step 4: Review instructions and permissions

Read the frontmatter and body before invocation. Check whether the skill permits model invocation, pre-approves tools, injects shell output, calls scripts, reads broad paths, or expects environment variables. Open supporting scripts rather than trusting their file names. Ensure secrets are read from an appropriate secure environment and are never committed into SKILL.md.

If the skill can deploy, publish, send messages, change infrastructure, or delete data, prefer an explicit user-invoked workflow. Installation does not remove the need for Claude Code's permission controls or human review.

Step 5: Confirm Claude Code discovers the Skill

Start Claude Code in the relevant project and use the current skills listing or slash-command menu to look for the installed name. Then test it in two ways:

1. invoke it directly with /skill-name;
2. ask a natural-language question that clearly matches its description, if automatic invocation is enabled.

Verify the response follows the intended procedure and can locate every referenced file. Also try a nearby unrelated request; the skill should not activate merely because it contains broad words such as “code” or “help.” Claude Code watches established skill directories for changes in current versions, but Anthropic notes that creating a new top-level skills directory after the session starts may require a restart.

Plugin-provided Skills

Plugins can distribute skills alongside agents, hooks, or MCP configuration. Install and manage these through the plugin's documented workflow rather than copying a nested folder out of its package. Plugin skills use a namespaced command such as /plugin-name:skill-name, which helps them coexist with personal and project skills.

Common installation mistakes

• Putting SKILL.md directly in .claude/skills/ instead of inside a named folder.
• Adding an extra wrapper directory, such as skill-name/skill-name/SKILL.md.
• Renaming the directory without updating the standard name field.
• Copying the instruction file but omitting referenced scripts or assets.
• Installing into the wrong project or starting Claude Code above an unrelated repository.
• Using malformed YAML or placing content before the opening frontmatter marker.
• Assuming a skill is trustworthy because it appears in a public collection.

If discovery fails, use the focused Claude Code skill storage guide to confirm scope, paths, and name collisions.

Updating or removing a Skill

Before updating, review the upstream diff and repeat your trigger and behavior tests. Do not replace a locally adapted skill blindly; preserve intentional project-specific changes or maintain a documented fork.

To remove a personal or project skill, delete that specific skill directory—not the entire .claude/skills tree. Current Claude Code releases detect changes inside known skill directories, although already-loaded instructions can remain in the active conversation. Start a fresh session when you need to verify the clean post-removal state. For a plugin-provided skill, disable or uninstall the plugin using its supported workflow.

PiSkill's Find Skills guide can help with discovery, while the source-linked skills directory lets you review individual publishers before installing anything.