The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external CSV and JSON data exported from financial systems, creating a potential surface for indirect prompt injection.
Ingestion points: scripts/import_tables.mjs reads user-provided local files for orders, invoices, and payments.
Boundary markers: All flagged anomalies are presented in a human-review queue at the #/anomalies route of the included AirApp, ensuring user oversight.
Capability inventory: The skill uses busabase-sdk for database operations and provides an execution plan for follow-ups. Actual execution (like sending emails) is intended to be performed by the agent through other dedicated skills after user approval.
Sanitization: Input data is normalized in content/kelly-audit-app/app/js/audit-model.js through functions like dateOnly and normalizeAmount.
[DATA_EXPOSURE]: The skill handles sensitive financial records including orders, invoices, and payments.
Evidence: scripts/import_tables.mjs reads local export files. The skill includes explicit instructions in SKILL.md to treat this data as sensitive business data and avoid committing export or environment files.
[COMMAND_EXECUTION]: The skill includes administrative scripts (import_tables.mjs, run_checks.mjs, execute_decisions.mjs) designed to be executed via Node.js to manage the audit workflow.
Evidence: These scripts are used for data ingestion and rule processing, following standard patterns for AI agent tooling.

kelly-audit