Skip to content
Founding Lifetime: every Pro skillfirst 100 members · $199 onceUpgrade
Lookbk
Sign In

Safety checks before running a skill

A skill is instructions and code your agent may run. Six checks to make first: its source, its commands, its network use and what it may do without asking.

For
Anyone about to run a skill they didn’t write
Written for
Claude CodeCodexCursorGitHub CopilotGemini CLIOpenCodeAny agent that reads SKILL.md folders
Time
4 min read · updated Oct 26, 2026

The short answer

  1. Treat a skill like software you install. Use skills from sources you trust, and read the files before you run one.
  2. Look at what it runs, what it fetches from the internet, and which tools it pre-approves.
  3. Keep your agent’s permission prompts on, and try a new skill in a throwaway folder first.
Jump to:Why a skill needs a lookKnow where it came fromRead SKILL.md and the scriptsSee what it fetches

Why a skill needs a look

A skill gives an agent new abilities through instructions and code. That is its value, and it is also the risk: a skill that does something other than what it says can direct the agent to call tools or run code you didn’t expect, and in the worst case leak data or reach parts of your computer. Anthropic’s guidance is to use skills only from sources you trust, and to audit any other one thoroughly. VS Code’s docs say to always review shared skills before using them.

Before you run a skill

Know where it came from

Prefer skills you wrote or got from a seller you trust. A Lookbk skill page names the open-source project the skill builds on and its license, and each download carries a NOTICE.md listing the third-party software and fonts it uses. The .lookbook/license.json file in a download names the account it was issued to.

Read SKILL.md and the scripts

Review every file in the folder, not only SKILL.md: the scripts, templates and anything else. Look for network calls, file access and commands that don’t match what the skill says it does.

Terminal
ls -R <skill-folder>
grep -rnE "curl|wget|fetch\(|https?://" <skill-folder>

In a Lookbk skill, the steps in SKILL.md run Node.js scripts from the skill folder and tools such as HyperFrames through npx, with the version pinned (for example npx --yes hyperframes@0.8.192 in Kinetic Title Drop). They also open Chrome or Chromium for checks and exports.

See what it fetches

Fetching from the internet is the risk to look at hardest. Content a skill pulls in might hold instructions of its own, and a trustworthy skill can turn bad if something it depends on changes later. Lookbk skills load fonts from Google Fonts and libraries such as GSAP or three.js from cdn.jsdelivr.net, and say so under “What you need”. Many give offline steps in their Troubleshooting section.

Check what it pre-approves

A skill’s header can list allowed-tools, the tools the agent may use without asking you. In Claude Code the grant lasts for the turn that runs the skill, and it applies to a project skill even in a folder you have never trusted, so read it before you start Claude Code in a repository that carries skills. GitHub warns to pre-approve shell or bash only for a skill whose scripts you have reviewed and fully trust.

Keep the permission prompts on

Your agent has its own controls. Leave them on, and use them to narrow what a skill can do.

Where each agent lets you control skills
AgentWhat it gives you
Claude CodeSkill(name) allow and deny rules in /permissions, and disable-model-invocation: true to stop automatic use.
Codexallow_implicit_invocation: false in a skill’s agents/openai.yaml stops automatic use. You can still call it with $.
Copilot CLIAsks before any tool the skill doesn’t list in allowed-tools.
Gemini CLIAsks before it activates a skill, showing its name, purpose and folder.
OpenCodePer-skill permissions in opencode.json: allow, ask or deny, with wildcards.

Try it somewhere safe first

Install it into a scratch repository with --project, run it there, and look at every file it made. A skill’s outputs belong in your working folder, so you can delete the whole folder if the result is wrong. Update, roll back and remove skills covers cleaning up.

Do

  • Read the scripts a skill ships before you run it.
  • Answer an agent’s permission prompt, not “yes to all”.
  • Stop and remove a skill whose steps ask for things the job doesn’t need, such as passwords or keys.

Don’t

  • Pre-approve shell or bash for a skill you haven’t read.
  • Install a skill from a link a stranger sent you.
  • Commit a skill to a public repository. Lookbk’s terms don’t allow sharing skill packages.

Common questions

Does installing a skill run it?
No. The installer downloads the files and writes them to a skills folder. The agent runs a skill’s steps only when you ask for the job, or when your request matches its description.
What does a Lookbk skill do on my computer?
The agent follows its SKILL.md: it writes files into your project, runs Node.js scripts, and calls tools such as HyperFrames, FFmpeg and Chrome that the skill’s page lists under what you need.
Can I stop an agent running a skill by itself?
Yes. Most agents let you limit a skill to manual use. See Write a description that makes a skill trigger for the settings.

Sources

What this guide says about other companies’ tools comes from their own documentation, checked on Oct 26, 2026.

Did this guide help?

Read next

All guides

Update, roll back and remove skills

Update an installed skill, go back to the version you had, and remove or switch off a skill in Claude Code, Codex, Cursor, Copilot, Gemini CLI or OpenCode.

Skill not working? Find your symptom

Fix a Lookbk skill by what you see: the agent can’t find it, it won’t trigger, the wrong skill runs, output looks off, a render fails or install stops.

Where skills live: folders for every agent

The folders each agent reads skills from, which copy wins when two share a name, and what cloud agents can see.

See a skill’s result before you install it.

Every skill page shows what it makes, what it needs and which agents can run it.

Browse skills See pricing