Turn mixed handwritten forms into reviewable JSON

Extract form fields and transcription while surfacing uncertainty for human confirmation.

Goal

Goal: Extract form fields and transcription while surfacing uncertainty for human confirmation.

Define a JSON field schema and null policy; send one authorized image as type: "image" with a precise instruction. [transcribe]
Ask for handwritten/printed transcription, field values, and uncertainty markers; forbid guessing ambiguous characters. [transcribe]
Parse the structured response and validate fields/types; route uncertain fields with source-image crops to a reviewer. [transcribe]

A reviewer can verify every field against the original image; invalid JSON and missing values are not silently accepted.

If handwriting is unclear, crop and retry per field or retain null; never overwrite the original scan.

Python 3.11+, Anthropic API key, a vision-capable model, de-identified scans, and a JSON validator.

Forms may contain personal or health data; use approved samples, restrict access/retention, and require human confirmation.

Transcription depends on image quality; schema validity does not prove field correctness.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT