Before you install an agent skill, read it

Before you install an agent skill or plugin, read it. It runs with your permissions, and most install guides tell you to pipe a script into your shell.

I cloned a handful of popular agent repos this week to read their install steps. The same patterns kept showing up.

1. 1A script piped straight into a shell. Free Claude Code, OfficeCLI and codebase-memory-mcp all offer a curl to sh or bash one-liner. Download the file and read it before you run it.
2. 2Plugins that run with your account's permissions. Ruflo's README says it plainly: its mods run with your account's permissions, so review their code before installing.
3. 3Installers that touch other tools. OfficeCLI's install step adds its skill to every coding agent it finds on your machine. That's convenient, and it's also a change to your setup you didn't make one tool at a time.
4. 4Keys in environment variables. Claude Context takes an OpenAI key and a Zilliz token on the command line. Use scoped keys with spend limits, not your main ones.

None of this means these projects are unsafe. It means the default way to install them asks for a lot of trust, and reading a SKILL.md and an install script takes five minutes.

What's your rule before you install something an agent will run?