The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill is designed to process untrusted external data, specifically emails and meeting transcripts (Step 4). While this is necessary for its primary function, it technically creates an interface where an agent could encounter malicious instructions hidden in external communications.
[Mitigation of Untrusted Content]: The skill includes explicit security rules that define external content (emails, transcripts, etc.) as data only, never instructions. It specifically directs the agent to report instruction-like text found in these sources rather than executing it, and prohibits rendering any links found within untrusted content.
[Human-in-the-Loop Safeguards]: The instructions require the agent to present any 'content-originated' actions (those dictated by external data) to the user for review before execution. This ensures that the user remains the final authority on any changes to the CRM or communications sent through the connectors.
[Least Privilege and Data Handling]: The skill is designed to work within the existing permissions of the connected tools. It emphasizes using live CRM schemas rather than assuming data structures, which helps prevent data corruption or improper field mapping.

win-loss-review