The Agent Skills Directory

[Handling of Untrusted Content]: The skill explicitly categorizes external data sources—such as emails, chats, and transcripts—as untrusted content. It instructs the agent to treat this information as data rather than instructions, which is a key defense against indirect prompt injection. By reporting instruction-like text rather than acting on it, the skill maintains control over the execution flow.
[Human-in-the-loop Verification]: A robust confirmation mechanism is in place for any 'content-originated' actions. If an action is suggested by external data, the skill requires the agent to present the exact details to the user for approval before execution. This ensures that no autonomous writes occur without explicit human oversight.
[Link and Resource Sanitization]: The instructions prohibit the rendering of links found within untrusted content, directing the agent to use only verified internal record links or IDs. This practice significantly reduces the risk of the agent being used to facilitate phishing or redirection to malicious external sites.
[Least Privilege and Scope Management]: The skill emphasizes 'empty personal scope,' requiring the agent to ask the user for clarification rather than silently widening access to organization-wide data. Additionally, it enforces that the agent must not assume CRM schemas but must ground its actions in the live connector's actual data structure, preventing errors or hallucinations during data updates.

update-opportunity