The Agent Skills Directory

[Ingestion of Untrusted External Content]: The skill is designed to process data from external sources such as emails, chat transcripts, and documents. This creates a potential surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent's context through connectors for email, chat, transcripts, and external documents as described in the Flow and 'Learn the user's writing voice' sections of SKILL.md.
Boundary markers: The skill includes clear instructional boundaries, labeling these sources as 'untrusted content' and specifying they should be treated as data, not instructions.
Capability inventory: The skill interacts with CRM systems, email clients, and calendars via connectors, and generates artifacts or 'Pages' for the user dashboard.
Sanitization: Instructions are present to identify and report instruction-like text within data and to avoid rendering links found within these untrusted sources.
[Processing of Sensitive Communication]: To learn a user's writing style, the skill analyzes sent messages from connected email accounts. Although this involves accessing private communications, the logic is limited to extracting stylistic patterns (such as tone, formality, and length) and explicitly forbids acting on content within the messages as instructions.

setup