The Agent Skills Directory

[Defensive Handling of Untrusted Data]: The skill explicitly categorizes content from emails, transcripts, and external documents as 'untrusted.' It instructs the agent to treat such content strictly as data rather than instructions, which is a key defense against indirect prompt injection.
[Human-in-the-Loop for Content-Originated Actions]: For actions where the recipient or content is dictated by external data (e.g., an email address found in a transcript), the skill mandates showing the exact details to the user for review before execution. This provides a significant safeguard against unauthorized data exfiltration or unintended messaging.
[Credential and Permission Management]: The skill adheres to the platform's native connector permissions and security models. It avoids hardcoded credentials and specifies that permissions are managed at the connector level, ensuring it cannot perform actions the user or admin has not explicitly allowed.
[Link and URL Safety]: There is a specific instruction to avoid rendering links found within untrusted content. Instead, the agent is directed to link directly to the verified record or thread ID, which helps prevent potential phishing or malicious redirections embedded in external data.
[Operational Transparency]: While the skill requests to 'work silently between tool calls' for efficiency, it balances this with instructions to show all proposed changes and evidence to the user, ensuring that updates to CRM records or document creation remain transparent and user-controlled.

close-plan