The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with a local binary at /opt/homebrew/bin/gog. It invokes this CLI to perform various operations on Google services.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it processes untrusted external data.
Ingestion points: It reads content from Gmail messages, Google Drive files, and Google Docs, which can contain instructions from third parties.
Boundary markers: The instructions explicitly mandate the use of the --wrap-untrusted flag when processing data that may contain untrusted content to prevent injection attacks.
Capability inventory: The agent can execute shell commands via the gog CLI, allowing it to read, search, and (with explicit intent) modify Google account data.
Sanitization: The skill relies on the CLI's --wrap-untrusted and --json flags to handle data safely, alongside instructions for the agent to use --readonly by default.

gog