Handwritten Table OCR Assistant

Problem

Handwritten tables are common in attendance sheets, sign-in logs, orders, expense claims, and vouchers, but the source data usually stays as images. Manual entry can drop names, shift columns, or make batch review hard. This skill targets the specific gap between a photographed table and a structured workbook: it converts one or more handwritten table images into validated JSON, then exports them into Excel for further processing.

How It Works

The skill combines multimodal recognition, schema validation, and an export script. For a single image, the agent reads the photo, applies a prompt template from references/prompt_templates.md, extracts the requested fields, and emits standard JSON. Ambiguous cells are marked as uncertain. For multiple images, recognition runs in parallel, each image produces an independent result, and progress is reported.

Core capabilities:
- Single-image extraction: capture names, dates, amounts, quantities, and other fields
- Batch processing: run multiple table images concurrently for multi-source datasets
- JSON validation: check required fields and structure against references/json_format.md
- Excel export: generate .xlsx files through scripts/json_to_excel.py
- Merge mode: align similar column names at ≥60% similarity, such as "Customer" and "Customer Name"

Export can produce:
1. Multiple worksheets: one per image, useful for tracing source files
2. One merged worksheet: normalized columns plus a Source Image column and summary stats

Limits

This skill is focused on recognition and structuring, not recalculation. It preserves the recognized raw content by default and does not perform business calculations; if a mixed column contains "formula = result", it can split it into "Formula" and "Result". Illegible handwriting, poor framing, or unclear field definitions can still affect accuracy, so scenario-based prompt templates should define field names, formats, and validation rules before batch export.