The Agent Skills Directory

[Untrusted Content Handling]: The skill implements a proactive security measure by explicitly classifying emails, chat logs, transcripts, and external documents as untrusted content. It instructs the agent to treat such data strictly as information and never as instructions, which effectively mitigates risks associated with indirect prompt injection.
[Manual Action Confirmation]: To ensure user oversight, the instructions mandate that any action dictated by untrusted content—such as identifying a recipient or target from an email—must be explicitly shown to the user for approval before execution.
[Scheduled Execution Constraints]: The skill includes constraints for scheduled or unattended runs, preventing untrusted content from introducing new actions into the workflow without human intervention.
[Secure Data Access]: The skill utilizes official platform connectors for accessing CRM and communication tools. It emphasizes using verified web links and forbids the use of inferred or guessed URLs, reducing the risk of data exposure or redirection to unauthorized domains.
[Indirect Prompt Injection Surface]: While the skill ingests data from untrusted sources (emails, transcripts) and possesses write capabilities (CRM updates, meeting scheduling), it provides clear boundary instructions and mandatory user confirmation steps to sanitize the impact of potentially malicious embedded content.

customer-health