The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data in the form of incoming email bodies. This creates a surface for Indirect Prompt Injection, where a sender might include instructions to influence the draft generation.
Ingestion points: The get_message_details function in references/gmail-api.md extracts email bodies for LLM processing.
Boundary markers: Not explicitly defined in the prompt assembly logic, though the skill uses separate analysis and drafting steps.
Capability inventory: The skill has the ability to send emails and create drafts via the gmail.send and gmail.compose API scopes.
Sanitization: The skill includes a robust check_sensitive_data function using regular expressions to detect and redact sensitive information like PII and credentials before a draft is finalized.
[EXTERNAL_DOWNLOADS]: The skill communicates with official Google API endpoints (googleapis.com) to manage email data. These are well-known, trusted services and do not pose a security risk in this context.
[SAFE]: The implementation strictly adheres to a human-in-the-loop (HITL) model, where the SKILL.md workflow explicitly mandates user approval (Step 5: User Approval) before any email is sent, preventing autonomous data exfiltration or unauthorized communication.

gmail-reply