A scribe for a language that's almost gone.

A family has a notebook. A grandmother wrote it in a regional dialect that nobody under fifty still speaks fluently. The OCR apps mangle it. Google Translate doesn’t recognize half the words. The dictionary it would belong to does not exist.

Heirloom is what happens when you stop trying to fix that and instead make a tool for the family to do it themselves.

How a page gets made

Upload a photo. The server downscales the image, normalizes the HEIC mess that iPhones produce, and sends it to Claude Opus 4.7 with a system prompt written like an instruction to a junior scribe. Claude returns a strict JSON object: a language guess, a transcription, a draft translation, and an explicit list of words it isn’t sure about.

That last field is the one that matters. Most AI tools smooth over their uncertainty. This one is told to flag it.

The elder reads the page on a phone. Any word looks wrong, they tap it, record themselves saying the word out loud, and the audio attaches to that character range in the transcript. The result is a shareable page that holds the scan, the corrected text, and the recordings. The URL is a 128 bit hex string so nobody stumbles onto a family’s artifact by accident.

What the model is doing, and isn’t

Claude is the scribe, not the author. The system prompt says so out loud. It’s also told that the audio, not the transcription, is the source of truth, because token level bounding boxes on handwritten regional script are about as reliable as you’d expect.

Translation is always labeled draft. The text Claude returns is a starting point for the elder to react to, not a finished output.

Built in ten hours at Kupfrian Hall for the Claude Builder Club Spring 2026 hackathon, Creative Flourishing track. The rate limiter is in process and resets on restart. Translation is always a draft. None of that is going to scale to a million families. It does scale to one.