The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-generated transaction notes from Venmo CSV exports, creating an inherent attack surface for indirect prompt injection if external text contains instructions designed to mislead the AI classifier.
Ingestion points: Venmo CSV transaction logs uploaded by the user in SKILL.md workflow steps 1 and 2.
Boundary markers: Absent. The instructions do not prescribe specific boundary delimiters or isolation techniques for the Note text content.
Capability inventory: Internal tools include transaction_search, categorize, anomaly_detect, and export_transactions for managing transaction records.
Sanitization: Absent. Heuristics are used for regex pattern matching, but explicit defensive prompt engineering or sanitization of data contents is not detailed.

venmo-reconciler