The Agent Skills Directory

[Untrusted Content Handling]: The skill includes robust instructions to treat external data—such as emails, chat transcripts, and external documents—strictly as data rather than instructions. It specifically mandates reporting any instruction-like text found within these sources without acting on it, which serves as a significant defense against indirect prompt injection.
[Controlled Action Execution]: There is a clear requirement for user confirmation before executing any actions that originate from external content. This ensures that the agent cannot be misled by untrusted data into performing unauthorized tasks, such as sending emails or updating records based on external input.
[Least Privilege and Data Integrity]: The skill is defined as read-only, explicitly stating that scenario math never touches actual records. This design prevents unintended data modifications during modeling. Additionally, it requires the agent to link directly to official record IDs rather than rendering potentially unsafe links found within external communications.
[Scope and Permission Management]: The instructions require the agent to clarify the intended data scope before proceeding, preventing it from accidentally accessing organizational-wide data when a personal scope was intended.

deal-slip-scenario