Amaranth Brown Bag Series

Transcribing handwritten documents

Tips and recommendations

Choose your tool based on your project needs
Use one LLM to check the work of another
Preview of agentic processing

Existing scholarship on HTR and AI

Several scholars are already working on this

The field’s response. David Joseph Wrisley and Estelle Guéville, Medieval Manuscripts and the Computational Humanities (2026), ebook at UNM Libraries
A plain guide to every option. Michael Schonhardt et al., Automated Text Recognition: Theory, Platforms, Best-Practices, ebook at UNM Libraries
LLMs compared with Transkribus. Mark Humphries, generativehistory.substack.com
Full transcription chats. Jonathan Seyfried, jonathanseyfried.net

Choosing the right tool

Match the tool to the needs of the project

First, the language, age, rarity, and style of the handwriting.
Second, the quality of the scan or photo, and its layout.
Third, the number of documents you need transcribed.
Fourth, the budget, time, and tech comfort you can bring to the project.

Then choose around 10 document images for a pilot test

Choosing the right tool

Try Transkribus live

transkribus.org

Choose a single image of a handwritten document with text in horizontal lines
Drag the image into the box on the Transkribus home page
Watch how it performs, with no additional setup required

Drag in an image and see what comes back. Need an example? Use Andrew Mauney’s diary.

Choosing the right tool

The usefulness of Transkribus depends on the script

If you have a document in Precalit Sanskrit or Old Nubian, you’re in luck! Someone else has created models for those.

For a Catalan document in late fourteenth-century Gothic Secretarial Hand, you are not lucky.

That said, a lot of scripts have existing models; almost all Latinate scripts from the nineteenth century will work on Transkribus.

Transkribus can be trained for a variety of layouts

Commercial LLMs

Train a model, or try a commercial LLM

If you have a script that is not yet available on Transkribus, you can train a new model

However, training a Transkribus model might not be worth the time, so instead you can try a commercial large language model

Generally, the AI chatbots will also work for Latinate scripts after 1900

Table comparing transcription error rates for Gemini, Claude, and ChatGPT. — Humphries and Leddy's error-rate comparison, reformatted by Claude

An example using Claude Fable 5.1

An unexpected success

Sometimes the LLM produces quite impressive results, as with this classical Japanese handwriting on a silk handkerchief from the 1940s.

Photograph of classical Japanese handwriting on a silk handkerchief from the 1940s. — Classical Japanese handwriting, silk handkerchief, 1940s

Checking the transcription

Have a second model check the first

Current recommendation

Use Gemini as the first LLM: Gemini currently leads in performance for image recognition.
Prompt Gemini: Upload the image file and specify the year and the script.
Then go to Claude: Paste Gemini's output into this prompt for Claude.

Prompt for the second LLM

I am going to paste a transcription of the same image done by another LLM. Please look it over and use it to consider parts of the transcription that were ambiguous or difficult for you. Then make adjustments to your transcription and provide it again in full. Here is the transcription...

It will take some time to check the end result for each document

11:00–13:30

Sometimes, this only works as a check if you actually read the two transcriptions against each other, not just accept whichever one finishes second. Other times, I found that the output from Claude was obviously superior with so much of the text that it wasn’t worth the time to check to see if Gemini had done better. I also prompted Claude to explain when it overrode Gemini in a significant way.

Sample checking prompt

I am going to paste a transcription of the same image done by another LLM. Please look it over and use it to consider parts of the transcription that were ambiguous or difficult for you. Then make adjustments to your transcription and provide it again in full. Here is the transcription...

Agentic AI

What agentic AI changes

Agentic AI means letting an AI application edit files directly on your computer
OpenClaw, the first popular agentic AI app, contained significant security vulnerabilities
Today’s frontier agentic tools, such as Claude Cowork, are considerably safer
The October 21 Reliable AI Brown Bag goes into agentic AI in more detail

Agentic AI

Batches instead of one by one

Agentic AI can transcribe a whole batch of images instead of you uploading them one at a time
Tried on document images from the Archive of the Crown of Aragon, with varying levels of success
Batch output can go straight into a CSV, ready to open in Excel or upload as a database on a website

Two examples of CSV files are available on Jonathan Seyfried’s website

Cautionary notes

Points to take with you

Image recognition drains usage credits quickly

There is a substantial time commitment for perfect transcriptions, increasingly viewed as a ‘verification tax’

Consider the use of AI for transcription as more of an axe than a scalpel

At the same time, be ready for the AI to surprise you with moments of high accuracy

More resources for HTR using AI

The webpage has the links and more detail

amaranth.unm.edu/reliable-ai

The presentation slides
Links to all the examples covered in the presentation
A report on 2026 HTR developments generated by Claude
Citations of current scholarship