The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from tax documents provided by the user in PDF or image formats.
Ingestion points: Document paths are processed via the import_data.py script and specialized sub-skills for OCR (SKILL.md, Step 1, 1.6, 1.8).
Boundary markers: The skill mandates "dual verification" (comparing two independent LLM extraction results) and explicitly requires user confirmation via AskUserQuestion for any discrepancies or manual entry needs.
Capability inventory: The agent can write state information to .shinkoku/progress/ files and perform database operations (CRU) on a local shinkoku.db file using the ledger.py script.
Sanitization: A sanity-check operation is executed after calculation to verify logical consistency and compliance with rounding and tax rules defined in the reference files, serving as a functional filter for anomalous or malicious data embedded in processed documents.

income-tax