The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data from YouTube (such as search results, activities, and comment threads). It proactively mitigates the risk of indirect prompt injection by instructing the agent to use the --wrap-untrusted flag and machine-readable formats to help the agent distinguish between data and instructions.
[DATA_EXFILTRATION]: The skill interacts with YouTube account data. To prevent accidental data exposure or unauthorized changes, it mandates the use of --readonly mode when mutations are not required, --dry-run for testing writes, and explicit manual confirmation before any data is modified or deleted.

gog-youtube