What to check before installing an AI skill
Check the source, permissions, runtimes and recovery plan before enabling a skill. Use a small first test, and treat automated scanning as a review aid rather than a safety guarantee.
A useful-looking skill is still something to review before you enable it. Go beyond the feature description: read its instructions and bundled files, understand what they would do in your environment, and start with a small test rather than your real workload.
1. Check the source and the exact download
Review the seller profile, description, version and change notes. Check that download links still lead to the source you expected and that the file is the skill you intended to install. For redistributed material, look for an original source and licensing information you can examine.
If a checksum is provided, comparing it with your local file can help establish whether your copy matches the referenced file. A match does not establish that the contents are safe. An official label or a popular listing is not a substitute for reviewing the contents.
2. Read the instructions as a request for access
Open SKILL.md, the README and bundled scripts before enabling the skill. Compare what they request with the job you want done:
- Which folders will it read, and where will it write? Could it overwrite or delete existing files?
- What data would leave your computer, to which destination, and for what purpose?
- Does it need API keys or other credentials? Are the requested permissions broader than the task needs?
- Does it install software, run external commands, create persistent settings or request administrator access?
- Do the instructions ask the agent to skip confirmation, conceal actions or disclose secrets?
If the access seems excessive, stop and ask the seller rather than granting it automatically. Before sharing everyday credentials, consider whether a separate test credential with limited permissions is available.
3. Check runtimes and dependencies before execution
Read the listed runtime requirements, including Python or Node.js versions where applicable. Check compatibility with your operating system and agent, and confirm the intended installation location. Do not run a suggested repair command simply because it appeared in an error message; first understand what it changes.
When dependencies are required, check the package names and where they come from. Treat placing a skill folder and running bundled code or an installer as separate decisions. The Windows and Mac installation guide covers the placement steps for supported agents.
4. Prepare a way back, then use a small test
Back up important files before testing. Save the previous version of any configuration you intend to change. Prefer a copied working folder and sample input without personal information. Your first run should not target production documents, a large file collection or a live service.
After the test, inspect more than the result: look at changed files, requests for network access or permissions, and any new settings. If the behavior is unexpected, disable the skill and investigate instead of expanding its access until the error disappears. Repeat the review when you replace it with an updated version.
5. Start with the free official audit skill
Before using a new AI skill, we recommend installing AI Skill Safety Audit, provided free by AI Skill Market. Use it as a review aid to inspect a downloaded skill folder or ZIP before enabling the candidate skill.
It reads Python, Node.js, Shell Script, cmd and PowerShell code, along with SKILL.md and README instructions, to look for indicators such as destructive operations or credential access. It does not execute the candidate code, make network requests or install dependencies. Running the scanner itself requires Python 3.10 or later.
Follow the installation guide to get started. Review the audit skill itself, too, and use its findings as leads for manual inspection.
Automated checks can produce false positives and miss problems. A finding is not proof of malicious intent, and no findings is not a safety guarantee. Treat unreadable files and areas outside the scan's coverage as unresolved work rather than a clean bill of health.
Before enabling the skill, ask yourself: can I explain what it does, what it can access, and how I would recover if it goes wrong? If not, leave it disabled until you have enough information to decide.