The Agent Skills Directory

Indirect Prompt Injection: The skill ingests untrusted data from external sources such as emails, calendar events, and meeting transcripts to update CRM records and draft follow-up communications.
Ingestion points: Processes content from external email threads, Zoom transcripts, and chat logs using the WebFetch and Read tools.
Boundary markers: Includes explicit instructions in SKILL.md under the 'Approval gates' section to ignore any commands found within the read data and to verify sensitive requests with the owner.
Capability inventory: The skill has permissions to write to CRM platforms (HubSpot, Salesforce, Monday.com), create contacts, and draft outreach messages.
Sanitization: Implements logic to summarize content and extract specific structured fields rather than logging raw, unfiltered data.
Dynamic Tool Invocation: The skill employs a discovery-based flow for Salesforce operations, identifying and calling specific operations at runtime based on natural language descriptions.
Pattern: Uses a three-step process (discover → describe → dispatch) to interact with the Salesforce API, where operation names are determined during execution.
Context: This dynamic approach is governed by instructions that require user-level access permissions, prohibit delete operations, and mandate explicit approval for any write actions.

crm-autopilot