The Agent Skills Directory

Untrusted Content Handling: The skill processes data from diverse external sources including CRM records, emails, chat transcripts, and uploaded documents. To mitigate potential risks associated with indirect prompt injection, it includes explicit instructions for the agent to treat this content strictly as data rather than instructions. It further requires that any instruction-like text found within these sources be reported to the user rather than acted upon.
Human-in-the-Loop Safeguards: For actions that originate from untrusted content—such as drafting messages or updating record fields—the skill mandates that the agent present the exact recipients, targets, and content to the user for review before execution. This practice ensures human oversight for operations that could otherwise be influenced by external data.
Least Privilege and Scope Management: The skill emphasizes adhering to the underlying connector permissions and prohibits the silent widening of access from personal to organization-wide scopes. It also requires explicit user confirmation for suggested changes that were not directly requested, supporting a transparent operational model.
Secure Link Handling: Instructions are provided to use only official web links or valid IDs from the connected tools rather than inferring hosts from email addresses or tenant names. Additionally, it specifies adding parameters to document links (e.g., SharePoint/OneDrive) to ensure they open in a browser environment rather than triggering automatic downloads.

route-lead