The Agent Skills Directory

[SAFE]: The skill implements a robust 'human-in-the-loop' architecture. Invoices are extracted into a structured format and stored in a private Busabase workspace where they must be reviewed and approved via a dedicated spreadsheet-style UI before any data is exported to CSV or JSON.
[SAFE]: Secret management is handled securely. The import and export scripts utilize environment variables (e.g., BUSABASE_API_KEY) for authentication, and the documentation provides explicit instructions against committing sensitive credential files to version control.
[SAFE]: The skill uses official and well-known dependencies, primarily the busabase-sdk and the Hono web framework, which are standard for the platform's AirApp ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of invoice files (PDFs, images). While this represents a theoretical injection surface, the risk is effectively mitigated by strict schema validation and the mandatory human review step that prevents raw, unvetted data from being automatically processed or exported.
Ingestion points: Local invoice files (PDF, JPG, Doc, XLS) identified in the skill workflow.
Boundary markers: Extracted data is strictly delimited by the JSON batch schema defined in references/invoice-batch-schema.md.
Capability inventory: The skill has the ability to write to the local filesystem (the exports/ directory) and perform authenticated writes to the Busabase database.
Sanitization: Data is normalized into specific field types (money, dates, status codes) and requires explicit human approval in the review UI.

kelly-invoice-sheet