Invoice Organization and Reimbursement Pack Creation

Problem

Many invoice tools treat processing as OCR, so tax IDs, amounts, and dates may be guessed. This skill treats invoice work as a structured data pipeline: originals are preserved, field provenance is traceable, and uncertain fields are marked needs_review: true instead of being silently corrected.

How It Works

It moves through input, extraction, normalization, validation, export, and reimbursement packaging:
- Scan ZIPs, folders, or single files, create a manifest, and avoid moving, renaming, or deleting source files by default.
- Prefer xml, PDF/OFD text, filenames, tables, and QR data. Use multimodal vision for images first; fall back to external OCR only when vision fails, key fields are missing, or the user requests OCR.
- Normalize dates to YYYY-MM-DD, keep amounts as two-decimal strings where possible, and preserve source_path and file_hash.
- Validate amount relationships, required fields, duplicate candidates, and abnormal dates; low-confidence records go to needs_review.csv.
- Export to /发票台账/ under the original folder by default, with Excel, CSV, normalized JSON, review, and duplicate files; highlight review and reimbursed rows in Excel.
- Filter unreimbursed, non-review, non-duplicate invoices by a request such as “500 yuan of dining invoices”, create /报销包/..., manifest, and notes, then ask whether to mark them reimbursed.

Boundaries

It fits office invoice collection, ledger preparation, and reimbursement pre-checks, not financial audit or tax determination. Missing fields are not fabricated; privacy fields such as tax IDs, addresses, and bank details are handled conservatively, and external OCR or remote processing requires user confirmation.