The Agent Skills Directory

Untrusted Data Handling: The skill processes external content like call transcripts, emails, and chat messages. It addresses the potential for indirect prompt injection by explicitly instructing the agent to treat such content as data rather than instructions, to report any instruction-like text without acting on it, and to avoid rendering links found within these sources.
User-Controlled Actions: Any action that originates from untrusted text (such as drafting an email or updating a record) is flagged as a "content-originated action." These require the agent to present the exact details to the user for approval before execution, maintaining human-in-the-loop oversight.
Capability and Permission Scope: The skill utilizes standard platform connectors for CRM and communication tools. It adheres to the permissions established by those connectors and avoids assuming or guessing resource locations, preferring verified links and schema-grounded data access.
Data Sanitization and Verification: The skill is configured to cite sources with plain-text labels and human-verifiable links, ensuring that the user can validate the information retrieved from external documents and transcripts.

customer-voice