The Agent Skills Directory

[SAFE]: The skill implements a robust 'Approval Boundary' policy across all sub-skills (setup, receipts, reconciliation, monthly close, tax prep). It explicitly requires human confirmation before posting financial data, changing account settings, or finalizing compliance work, which mitigates the risk of autonomous unauthorized actions.
[COMMAND_EXECUTION]: The skill includes several Python scripts (e.g., receipt_summary.py, category_review_summary.py, reconciliation_summary.py). These scripts are used for deterministic data summarization (counting records and statuses) and rely solely on Python's standard libraries (json, sys, collections, pathlib). They do not perform any dangerous operations like shell execution or network requests.
[DATA_EXFILTRATION]: Although the skill handles sensitive financial data and requests the WebFetch tool, there is no evidence of malicious data exfiltration. The instructions focus on using the user's existing tools (Gmail, Outlook, accounting software) and prioritize high-provenance local evidence over remote sources.
[EXTERNAL_DOWNLOADS]: The skill mentions and provides links to receiptor.ai (the author's service), agentskills.io (the specification it follows), and well-known accounting tools (QuickBooks, Xero). These references are contextually appropriate and non-malicious.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as receipts, invoices, and bank statements. While this presents an inherent indirect prompt injection surface, the skill implements a 'semi-automated' execution mode with mandatory human review for all posting and finalization steps, effectively managing this risk.

bookkeeping