Installing Skills Safely

Installing a skill isn't like flipping on a harmless feature. It's closer to approving software.

Skills can have real access, real permissions, and real consequences. If your agent is going to install something that changes what it can do, the right mental model is software supply chain, not convenience setting. This isn't paranoia — it's basic operational hygiene.

The safe workflow is simple enough to remember.

Start from a filtered source. A broad marketplace is useful for discovery but it's not a recommendation. Safer starting points — curated lists that have already done some filtering — exist for a reason. They narrow the field before you begin evaluating.

Tell your agent to inspect the skill. What does it actually do? What access does it need? Does the author and source make sense? Is it asking for more than the job requires? This is where your agent does the first layer of practical skepticism.

Use the dedicated vetting step. The point of this final pass is to inspect actual code and permissions with a sharper lens before anything gets installed. Not vibes — a real check.

Only after those three steps should installation happen.

Why be this careful? Because the easy mistake isn't always an obviously bad tool. Sometimes it's a confusingly similar one.

One of the most useful cautionary examples in the whole system: two skills can share nearly the same name while coming from different authors and doing very different things. If someone picks based on surface similarity instead of checking the source and reviewing what it actually does, they can end up installing the wrong thing entirely. You don't need a dramatic attack story to justify caution. A mundane wrong-author install is enough.

This is why the right request to your agent isn't "can you just install this?" It's "please vet and install this skill." That small difference signals that the install is an approval process, not an impulse purchase.

You don't need to become the code reviewer yourself. Your role is to insist on the process. Your agent does the inspection and only proceeds when the result is clean enough to trust.

Broad marketplace for discovery. Curated source for a safer shortlist. Agent review for common-sense inspection. Dedicated vetting for the final technical check. Then install.

That may sound slower than clicking yes and moving on. Good. Speed isn't the point here. The point is to avoid importing risk just because a capability sounds useful in the moment. Handled well, skills become an advantage. The team gets new capability, and keeps the discipline that makes those capabilities worth having.