The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted customer data from multiple external channels (Email, WhatsApp, WeChat, etc.), which establishes an indirect prompt injection surface.
Ingestion points: Customer messages enter the agent context through the collection workflow and are stored in the tickets and messages bases, as processed in content/kelly-support-app/app/js/providers/busabase-provider.js.
Boundary markers: The application maintains a clear separation between customer transcripts and agent suggested replies in the AirApp UI (app.js), using structured data fields to prevent accidental instruction execution.
Capability inventory: The skill possesses capabilities to send replies, escalate tickets, and process refunds via scripts/execute_decisions.mjs, all of which are gated by status checks.
Sanitization: The runQualityGate function in support-model.js performs automated policy checks to ensure drafts cite valid knowledge base articles and do not contain unauthorized financial or service commitments.

kelly-support