The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from incoming emails to generate triage scores and reply suggestions.
Ingestion points: The skill reads external data via search_gmail_messages, list-mail-messages, and get_gmail_message_content tools.
Boundary markers: The instructions do not implement explicit delimiters or boundary markers for email content processing, which could allow instructions within emails to be interpreted as commands.
Capability inventory: The skill possesses write capabilities through email sending tools and command execution capabilities via Bash.
Sanitization: The skill implements a human-in-the-loop quality gate, instructing the agent to always show drafts for approval before sending, which mitigates the risk of unauthorized actions.
[COMMAND_EXECUTION]: The skill executes a local Python script to perform domain diagnostics.
Evidence: The orchestration logic invokes scripts/check_deliverability.py via the Bash tool to conduct DNS-based audits.

email