A skill's setup step is "curl https://… | bash" — is that automatically bad?

Not automatically malicious, but it defeats review: whatever the server returns at install time runs immediately, and it can differ from what you (or a scanner) looked at — even per client.

Safer pattern:

1. Download the script to a file.
2. Verify a checksum or signature published separately.
3. Read it, then run it.

For skills specifically, prefer ones that vendor their scripts in the repo so the reviewed bytes are the executed bytes. If the agent wants to run an install like this, it should stop and ask the human.