The Agent Skills Directory

[SAFE]: The skill implements a robust security model for email operations by requiring a 'Preview, show, confirm, send' workflow. The agent is strictly instructed to use the --dry-run flag to generate an exact copy of the intended email, display it to the user for review, and wait for a manual go-ahead before final delivery.
[INDIRECT_PROMPT_INJECTION]: The skill proactively mitigates prompt injection by treating all email content as untrusted and potentially malicious.
Ingestion points: The agent interacts with external, potentially attacker-controlled data when reading existing email threads for replies or forwards.
Boundary markers: The instructions provide a high-level security boundary by explicitly stating that message content is attacker-controlled and should not be used for instruction following.
Capability inventory: The skill allows reading local files for bodies or attachments and transmitting them via the gmail CLI tool.
Sanitization: Security is enforced through mandatory human verification of all generated email bytes and specific logic to report suspicious content to the user rather than executing requests embedded in emails.

gmail-send