Skip to content
MeloSkill

Guide

How to evaluate a skill before installing

A skill is instructions an agent may follow, including commands and scripts in the same folder. Treat it like running someone else’s code even when the file is markdown.

Installing a skill means your agent will read that author’s playbook on matching tasks. That can include shell one-liners, file writes, and helper scripts. MeloSkill does not execute any of it at build time. The listing is a reading aid.

What our badges mean

Each skill page has a safety line: Pass, Review, or Risk. It is a regex pass over the SKILL.md text. We look for secret-file paths, curl | sh patterns, pastebin-style dump sites, and “ignore previous instructions” phrasing. One hit is Review; two or more is Risk.

Pass does not mean safe. A skill can be polite markdown and still tell the agent to rewrite production CSS, commit secrets, or fetch a URL we have never seen. Details of the scan are on Methodology.

  1. 01 · Read the file

    Open SKILL.md on the listing or on GitHub. Confirm it does the job you asked for. If the description says “frontend” and the body is a marketing workflow, skip it — that mismatch is common in catch-all catalogs.

  2. 02 · Last change to this file, not stars

    The date on the listing is the last commit that touched this SKILL.md when GitHub gives it to us. A dead file in a 100k-star monorepo is still dead. Stars count the repository.

  3. 03 · Safety line

    Read Pass / Review / Risk and the flag text. Review is not a veto; it means go look at the matching paragraph. Risk means do not install unless you have read the file and accept it.

  4. 04 · Scripts

    If the listing says the folder includes .py, .sh, or .js, open those files on GitHub before you install. We never run them. An agent that follows the skill might.

  5. 05 · License

    Prefer a license you can live with. Unspecified licenses can appear in the index. They stay off task comparison pages and the homepage unless they come from Anthropic, Vercel Labs, or CloudAI-X and GitHub still has not published a license file — those listings still say Unspecified.

Red flags in the body

  • Asks the agent to read .env, SSH keys, or cookie stores “for context.”
  • Pipes a remote URL into a shell.
  • Posts repo contents to a webhook, pastebin, or tunnel host.
  • Tells the model to ignore its system prompt or hide actions from you.

A skill can be dangerous without any of those strings. The other check is intent: does this file teach layout, or does it teach the agent to send the working tree somewhere else?

What we will not claim

We do not run skills in a sandbox. We do not give letter grades. We do not invent install counts. If a listing looks wrong, email support@meloskill.com with the URL and the GitHub path.

Also