reading-invoice

Invoice image reading

A skill that reads invoice images and returns them as structured data.

For PDF files

The file is PDF (.pdf), attempt text extraction before image OCR.

1. shinkoku pdf extract-text --file-path <path> Run
2. If the extracted text contains the necessary information (amount, date, billing source, etc.), generate structured data from the text.
3. If the text is insufficient (scanned PDF, etc.), convert to PNG with shinkoku pdf to-image --file-path <path> --output-dir <dir> and proceed to the image reading flow below.

Image Reading Method

Recommended: Dual verification (2 parallel contexts)

To improve accuracy, read the same image in two independent contexts in parallel and compare the results.

1. Run two independent readings: In an environment where subagents are available, launch two subagents in parallel, each reading the image independently. Pass the following "Basic Rules" and "Output Format" to each subagent as a prompt, and specify the image file path.

2. Result comparison: Compare the key fields (amount, etc.) from both reading results.

3. If they match: Adopt them as is. Report that "the results matched in two independent readings."

4. If they do not match: Present the original image path and both results to the user, and ask them to choose the correct one:

   • Clearly indicate the fields with discrepancies
   • Present three choices: Adopt A / Adopt B / Manual input

Fallback (when subagents are not supported)

In an environment where subagents are unavailable, read using the following procedure:

1. Read the image file directly with the Read tool
2. Extract data according to the following "Basic Rules" and "Output Format"
3. Present the extraction results to the user andalways ask them to visually verify

⚠ Because dual verification is unavailable, always ask the user to visually verify.

Basic Rules

• Read image files with the Read tool (Claude Vision automatically recognizes images)
• Always return amounts as int (integer in yen). Remove commas and "円"
• Return dates in YYYY-MM-DD format
• Convert Japanese era names to Gregorian years (Reiwa 7 → 2025, Reiwa 6 → 2024, Heisei 31 → 2019)
• For fields that cannot be read, use UNKNOWN (string) or 0 (amount)
• If multiple files are passed, process all of them in order and return them together

Output Format

Read the image and return it in the following format:

---INVOICE_DATA---
vendor: 請求元名
invoice_number: 請求書番号
invoice_registration_number: 適格請求書発行事業者番号（T+13桁）
date: YYYY-MM-DD
total_amount: 請求金額合計（int）
tax_amount: 消費税額（int）
items:
  - description: 品目・サービス名
    amount: 金額（int）
    quantity: 数量（int）
    tax_rate: 税率（10 or 8）
---END---

Extraction points

• Extract the total invoice amount (tax included) with the highest priority
• Check the consumption tax amount (distinguish between the 10% and 8% reduced tax rate categories if present)
• Check for the presence of an invoice number (T + 13-digit qualified invoice issuer registration number)
• If there is an invoice number, extract it
• Extract the name of the vendor (billing source)
• Use the billing date for the date (if the issue date and billing date differ, prioritize the billing date)
• Extract line items to the extent they can be read (if unknown, leave items empty)

Processing multiple files

When multiple file paths are specified:

1. Get the file list with the Glob tool (if a pattern is specified)
2. Read each file in order with the Read tool
3. Return the results of all files together (write the file name before each result)

## file1.jpg
---INVOICE_DATA---
...
---END---

## file2.jpg
---INVOICE_DATA---
...
---END---