ScriptAIX

ScriptAIX is a small Python tool that turns scanned handwritten documents into readable, searchable text. You point it at a folder of scans or PDFs, and it sends each page to Claude, Anthropic's large language model, together with a carefully tuned expert prompt. The result is a diplomatic transcription: the original spelling, punctuation and line breaks are preserved, nothing is modernised, nothing is guessed. Every page ends up as a Markdown file, and an interactive viewer shows the scan and the transcription side by side.

It started with my genealogy research. The archives of Zeeuws-Vlaanderen and East Flanders are full of documents I wanted to read: 19th-century memories van successie (estate inventories), notarial deeds, 17th-century marriage registers, Flemish parish registers written in church Latin, private letters from the 1600s, and German family letters in Kurrent and Sütterlin script.

Reading these takes real palaeographic skill, and even with practice it is slow, tiring work. Modern multimodal AI models turned out to be remarkably good at it, provided you give them precise instructions about the document type, the region, the period and the transcription conventions. ScriptAIX packages exactly that into a repeatable workflow: scan in, faithful transcription out, with every uncertainty honestly marked.

ScriptAIX is a single Python script driven from the command line. It accepts JPG, PNG, WebP, multi-page TIFF and PDF files (with three interchangeable PDF backends), fixes image orientation, scales pages to the optimal resolution and sends them to the Anthropic API. Pages are loaded lazily, so even large PDF volumes do not fill up memory. Failed pages are retried automatically, already transcribed pages are skipped on the next run, and duplicate file names in different subfolders are resolved automatically.

Each transcription is saved as a Markdown file with a metadata header: source file, date, model, prompt preset and token usage. A run ends with a cost summary, so there are never surprises on the API bill.

The quality stands and falls with the prompt. ScriptAIX ships with a library of presets, each written for one document type and one script tradition:

• dutch-19th: 19th-century Dutch administrative handwriting
• memories-van-successie: Dutch estate inventories, with attention to heirs, assets and amounts
• ondertrouw-zeeuws-vlaanderen-17e-eeuw: 17th-century marriage and banns registers from Zeeland
• assendelft: early modern Dutch private correspondence (ca. 1600 to 1800), including text written sideways in the margins
• kerklatijn-meetjesland-18e: Flemish parish registers in church Latin
• suetterlin: German Kurrent and Sütterlin script
• kirchenbuch: German, Latin and Dutch church books
• general: a sensible default for everything else

All presets enforce the same conventions: uncertain readings are marked [?], illegible passages [onleesbaar], marginal notes [kantlijn: ...], strikethrough and insertions are preserved, and the model may append a short remarks block when something about the page is worth noting. Custom prompts can be added as simple JSON, no code changes needed.

Every run generates viewer.html: scan on the left, rendered transcription on the right. It offers zoom and pan on the scan, full-text search across all pages, colour-coded badges for uncertain and illegible passages, highlighted marginal notes and remarks, and metadata pills showing date, model, prompt and token count for each page.

The viewer doubles as an editor. Started with --serve, a local web server (bound to localhost only) lets me proofread and correct transcriptions directly in the browser. Saving writes the .md file and keeps a backup, a restore button undoes mistakes, and a reload button picks up pages transcribed while the viewer is open. The transcription data lives in a separate scriptaix.json, ready for other tools to consume.

For bigger collections there are three cost and throughput features. A dry run (--dry-run) lists all pending pages and estimates the cost before anything is sent, using exact token counting on sample pages. Batch mode (--batch) routes the whole run through Anthropic's Message Batches API at half price; the batch state is saved on disk, so the wait can be interrupted and resumed at any time without paying twice. And parallel mode (--threads) speeds up interactive runs, with the first page run separately to warm up the prompt cache.

ScriptAIX does not replace the palaeographer, it replaces the first draft. The model misreads names, numbers and faded passages just like a human does, which is why every uncertain reading is marked and why the viewer makes proofreading against the original scan as painless as possible. For my purposes, checking a good draft is many times faster than deciphering from scratch, and that is the entire point.

ScriptAIX is a personal tool and not (yet) published. If you are curious about it, email me.