This extracts structured data from Japanese withholding tax certificate images (源泉徴収票), the kind you need for tax filing. It handles both scanned PDFs and image files, pulling out payment amounts, withheld tax, social insurance, and insurance deduction breakdowns. The interesting part is the dual verification approach: when sub-agents are available, it reads the same image in two independent contexts and compares results, asking you to resolve conflicts if they don't match. Otherwise it falls back to single-pass OCR with mandatory manual verification. Converts Japanese era dates to standard format and outputs everything as structured integers. Useful if you're building Japanese tax prep tools or just tired of manually transcribing these forms.
npx skills add https://github.com/kazukinagata/shinkoku --skill reading-withholding