Read Before You Run: The 5-Minute Skill Audit
The skills ecosystem has a property worth saying plainly: anyone can publish a SKILL.md, and star counts measure attention, not correctness. There is no gatekeeper, and there should not be. But that means the trust decision is yours, per skill, every time. The good news: a real audit takes about five minutes, because skills are deliberately transparent. Everything the skill will do is in files you can read.
The audit
- Read the whole SKILL.md, the actual instruction file, not the README. You are reading what your agent will be told to do. Anything surprising (an instruction that touches files outside its stated job, phrasing that tells the agent to bypass checks or hide actions) is a stop, not a shrug.
- Map the trigger surface. From the description alone, list what requests this skill will catch. Too broad, and it will hijack requests that were never meant for it; that is a functionality risk even when intent is good.
- Open every script. Bundled code is code your agent may execute. Check what each script reads, writes, and, especially, whether it makes network calls or touches credentials. A helper that phones home is something you want to know about before it runs, not after.
- Hunt the assumptions. Hard-coded paths, expected environment variables, services presumed running. Undeclared assumptions are where "it worked for the author" quietly becomes "it broke on your machine."
- Record source and version. One line noting where it came from and which revision you reviewed. When the skill updates, that line tells you a re-read is due. An updated skill is a new skill.
What a trustworthy skill looks like
- A description that opens "Use when…" and names its own boundaries.
- A body that states its prerequisites instead of hiding them.
- Scripts that are small, dependency-free, and referenced explicitly from SKILL.md.
- Nothing in the folder that SKILL.md does not mention; every file is part of the contract.
Notice these are exactly the habits good authoring produces. The audit is symmetrical: what you check as an installer is what you owe as a publisher.
This is also our submission bar
When you submit a skill to this community, this audit is the review it gets. Skills that pass cleanly (readable, scoped, honest about their assumptions) are the ones that get featured. Which means the checklist above doubles as the pre-submission checklist: run it against your own skill before you send it.