The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines patterns for processing untrusted user-generated content from support tickets and knowledge base feedback.
Ingestion points: The triageTicket function in SKILL.md ingests ticket.subject and ticket.description. The rateArticle function ingests feedback strings.
Boundary markers: The provided code snippets do not include explicit boundary markers or instructions to the LLM to ignore embedded instructions within the user input.
Capability inventory: The skill demonstrates capabilities for database modification (db.articleFeedback.create, db.article.update, db.customerHealth.upsert) and automated notifications (notifyCSM). A successful injection could theoretically influence routing or health scoring logic.
Sanitization: While SQL injection is mitigated using Prisma's parameterized queries (Prisma.sql), there is no evidence of sanitization for text passed to AI classifiers like classifyTicket.

customer-success