Bill Classification Assistant

Problem

Personal bills often arrive as scattered Excel, CSV, and image files, with inconsistent headers, sheet layouts, duplicate rows, and export templates. This skill breaks bill cleanup into auditable local steps instead of pushing the entire workbook into the model context.

How It Works

The workflow first confirms the target template as icost or cookie, defaults currency to CNY, and creates a fresh run directory with workflow-config.json. It then runs deterministic stages:

• Use inspect_bill_file.py to inspect file type, encoding, sheets, row count, and preview
• For multi-sheet Excel files, ask the user to select transaction sheets; the agent does not guess valid sheets
• Run structure-detection on confirmed sheets to locate headers, transaction start rows, and key columns
• Chunk complete records with chunk_bill_rows.py, then normalize with normalize_table_chunks.py into billList, skippedItems, and warnings
• Recognize image bills with a vision-capable model while preserving source image metadata
• Validate, merge, deduplicate by datetime + price + type, export CSV, and generate an audit report

Boundaries

This skill is for local bill cleanup and does not rely on web services, login, Supabase, Mastra, or project code. It does not perform FX conversion, subcategories, transfer offsetting, or XLSX export. Full raw chunks are not retained by default unless privacy settings allow it.