The Agent Skills Directory

Untrusted Content Handling: The skill includes explicit instructions to treat all external data (emails, chats, transcripts) as data rather than instructions. It specifically directs the agent to report any instruction-like text rather than acting on it, which is a robust defense against indirect prompt injection.
Data Exfiltration Safeguards: There are clear restrictions against rendering links found within untrusted content. Instead, the skill mandates linking only to records via verified IDs or provided connector links, reducing the risk of phishing or malicious redirects.
Human-in-the-Loop Requirements: The instructions require explicit user confirmation for any 'content-originated' actions—those where the recipient, target, or message body is dictated by untrusted external text. This ensures that the user maintains control over critical operations like sending emails or booking meetings.
Least Privilege and Scope: The skill enforces an 'empty personal scope' policy, requiring the agent to ask the user for clarification rather than silently widening access to organization-wide data.
Secure Tool Usage: It uses standard calendar and CRM connectors and relies on the platform's native permissions and safety filters rather than attempting to bypass them.

schedule-meeting