The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute provided Python scripts (bankflow_precheck.py, bankflow_merge.py, and bankflow_reconcile.py) to analyze financial documents and generate reconciliation reports.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its 'AI Review' workflow. In this process, the agent reads transaction summaries and descriptions from user-provided files (exported via ai_unmatched.json). Malicious strings within these transaction records could attempt to influence the AI's matching logic or decision-making.
Ingestion points: User-provided bank statements and accounting ledgers are read by scripts/bankflow_reconcile.py.
Boundary markers: The AI instructions for the subagent focus on output formatting but do not implement specific delimiters or 'ignore' instructions for the raw transaction data being processed.
Capability inventory: The agent can read and write local files and execute specific Python scripts; it lacks network access or arbitrary system command capabilities.
Sanitization: Data is cleaned to improve matching accuracy (e.g., removing whitespace and brackets) but is not sanitized to detect or prevent prompt injection patterns.
Mitigation: The reconciliation script (bankflow_reconcile.py) performs a mathematical validation ('hard check') ensuring that all matched records align within a 1.0 currency unit tolerance, which protects financial accuracy regardless of the AI's matching suggestions.

bank-flow-reconciliation