The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (Gmail threads and calendar events) to extract commitments and identify "open loops."
Ingestion points: Gmail thread content and calendar event data are ingested via the gbrain sync command.
Boundary markers: The instructions do not specify any markers or instructions to isolate the ingested email text from the agent's prompt context.
Capability inventory: The skill utilizes shell command execution through the gbrain CLI and network synchronization capabilities.
Sanitization: No sanitization or filtering steps are documented for the ingested content before commitment extraction occurs.
[COMMAND_EXECUTION]: The skill relies on executing shell commands via the gbrain CLI. It supports a --token-command flag in the gbrain sources add command, which allows for the execution of an external command to retrieve authentication tokens at runtime.
[CREDENTIALS_UNSAFE]: The skill provides setup instructions for Google OAuth credentials (client secrets and JSON files). It includes explicit security warnings to avoid passing these secrets in plain chat text and recommends using environment variables or saving them to local files with restricted (0600) permissions.

google-loops