The Agent Skills Directory

Indirect Prompt Injection Considerations: The skill processes external data like transcripts and notes which may contain embedded instructions. To address this, it includes several defensive patterns: (1) Ingestion Analysis: Untrusted data enters the context from Gong, Google Drive, SharePoint, or user pastes (referenced in SKILL.md). (2) Boundary and Instruction Logic: The skill explicitly instructs the agent to treat these sources as data only, ignore any instruction-like text found within them, and report such occurrences to the user. (3) Capability Management: While the skill can perform operations like CRM updates and email drafting, these actions are gated by user interaction. (4) Sanitization and Verification: It implements human-in-the-loop checkpoints, requiring user approval for any content-originated action, and ensures recipients are resolved via authenticated system metadata rather than potentially manipulated text.
Human-in-the-Loop Validation: For sensitive actions like updating CRM records or sending emails, the skill mandates that all proposals be shown to the user with specific citations and source links. This ensures that the user maintains oversight of all changes and outgoing communications.
Secure Target Resolution: By resolving recipients and target records using trusted system data (such as CRM contacts or calendar events) instead of extracting them from untrusted transcript content, the skill reduces the risk of misdirected data or unauthorized actions.

call-summary