The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data, creating an indirect prompt injection surface. This is mitigated by explicit instructions (Rule 3) to treat email bodies as hostile and ignore any embedded instructions. Ingestion points: Gmail connector reading email message bodies. Boundary markers: Detailed instructions to prioritize user commands over data content. Capability inventory: Gmail and Xero (read), Airtable and Google Sheets (write). Sanitization: Mandatory human confirmation step for all write operations.
[DATA_EXFILTRATION]: Rule 4 prevents unauthorized data transfer by prohibiting the auto-opening of links or images found in emails, which are identified as potential exfiltration paths.
[COMMAND_EXECUTION]: The skill enforces strict user oversight via Rule 2, which requires explicit approval before committing any data to external trackers like Airtable.

expense-wrangler