How to Install and Use Agent Skills — Getting Started

Guide and linked CLI documentation reviewed 1 November 2026. Agent setup and package requirements can change.

What you are installing

A skill is a folder with a SKILL.md instruction file and, sometimes, scripts, templates or references. Your agent uses that material to follow a particular workflow. AI Craft Pad helps you choose a source and understand the evidence. You install and use it in your own agent environment.

1. Choose a task and source

Start with one outcome: review a page, investigate a reproducible bug, or prepare an editable document. Search for that task, then select your agent. The agent filter reflects source compatibility evidence; it does not mean every listed workflow has passed a test in your client.

Open a result and read its purpose, tradeoffs, source facts and installation section. Follow the pinned source-folder link and inspect supporting files. Check the license, required services and what data the workflow sends elsewhere. Unknown means that evidence is missing. A high Skill Score, a first-party source or a clean static scan does not guarantee runtime safety. Understand the score and review limits.

2. Prepare your environment

You need a skills-compatible agent, a terminal in a small test project, and any dependencies described by the source. A regular chat window may not provide a terminal or load local skill folders. Follow your chosen agent's own setup instructions.

The current skills CLI documentation requires Node.js 22.20 or later. Check your terminal before installing:

node --version
npm --version

If either command is missing, use the official Node.js download instructions, reopen the terminal and check again. Run the first trial on disposable sample files or a separate branch so that you can inspect changes. An agent account, external API or cloud service can have its own charges; a catalog listing does not include those services.

3. Install one skill

Return to the skill detail page and use Copy command under Install from the source. Paste that command into the terminal for your test project. Read the command and installer prompts before approving them. Our catalog commands identify a source snapshot; the CLI package invoked by npx is a separate dependency.

Choose the intended agent and keep the first installation in the project. The CLI uses project scope by default; global scope shares a skill across projects. Select one skill rather than an entire repository. If your system cannot create links, consult the CLI's documented --copy option.

Check the installed list from the same project directory:

npx skills list

A listed skill confirms that the installer recorded it. It does not prove that your agent loaded it or that its task succeeded. Check your client's available-skills view or loading instructions, and reopen the session if that client requires it. Read the CLI installation and scope reference.

4. Try a small task

For a browser-testing skill, supply one existing page and ask for an assessment before any edits. Replace the bracketed text with the exact installed skill name:

Use [installed skill name] to review this one page.
Check keyboard navigation, text readability and mobile overflow.
Do not change files yet. List the affected elements, the evidence
for each finding and the smallest useful correction.
If you cannot inspect a required state, say what you could not check.

Provide the page or project files your agent actually needs. A skill does not give the agent missing browser access, dependencies or permissions. For a document skill, use a sample document and inspect the rendered pages. For debugging, provide a failing command, expected behavior and a reproducible case. Explore Skill Kits when you need complementary roles.

5. Check the result

Task: did the output answer the specific request?
Evidence: can you reproduce a reported defect or inspect the generated artifact?
Boundaries: were file changes, commands and external requests consistent with the permissions you gave?
Limitations: did the agent identify what it could not verify?

For the page-review example, visit the named elements, use the keyboard and inspect the stated mobile width. Confirm findings against the actual page before accepting a correction. Keep the source commit, environment and outcome with your project notes. An agent saying that it used a skill is not independent evidence of a correct result.

Update or remove it deliberately

These commands inspect available updates and open interactive removal respectively:

npx skills check
npx skills remove

Review changed source instructions before applying updates with npx skills update. A command containing a commit identifies a particular snapshot; a daily catalog refresh does not update files already installed on your computer. Confirm the source and version you intend to use. Removal changes installed skills, not edits or external actions already performed by an agent.

If something does not work

The command is not recognized: Check Node.js and npm in this terminal, then reopen it after installation.
The skill is installed but not used: Confirm the project directory, selected agent and installation scope. Use the client's skill-loading documentation and name the skill explicitly in a small matching task.
The installer reports no skills: Check the repository and skill name against the linked source. The CLI requires valid name and description metadata in SKILL.md. Report a broken catalog command through Contact with the catalog URL and error text, excluding credentials.
The agent asks for an account, API key or payment: Re-read the source requirements and choose whether that service fits your task. Authorize access through the service's own supported flow.
The result is poor: Reduce the task, supply a concrete example and inspect the stated limitations. Compare another source by its workflow and evidence.

Sources and next steps

This is an AI Craft Pad walkthrough. It is based on the Agent Skills format overview, the skills CLI requirements and its troubleshooting reference. The CLI documents anonymous usage telemetry and its opt-out settings; review those settings if they matter to your environment.

Choose a reviewed starting point → · Browse by agent → · Compare workflows →

Find a skill. Try one real task.