Early Modern Studies

Fable 5 Transcription Kit

The Fable Transcription Kit allows you to upload images of manuscripts and receive a transcription. The Fable Kit was tested in a blind protocol where agents were given a manuscript page, some prompting, and asked to produce a transcription. Across 15 experimental runs that included five different hands ranging from legible to very challenging, the Fable Kit beat the current benchmark of a Character Error Rate (CER) under 5%. The experimental runs averaged a CER of 3.16%. All CERs were calculated by an independent review script, so the transcribing agent had no knowledge of its score. Likewise, the transcribing agent had no knowledge of the ‘correct’ transcription.

Currently, Fable is only available to Claude Pro, Max and Enterprise Subscribers, and via the API where you can purchase credits. During the experiment, the cost per transcription averaged $7.96 in API pricing with prompt caching.

If you use a model other than Fable, then you will not get the same results and will likely receive a transcription with a much higher CER. The next toolkit will be tailored to a model that does not require a subscription.

1) For subscribers

Once you’ve downloaded the toolkit, here’s how to use it if you have a Claude subscriber account or plan to purchase one:

  1. Go to claude.ai.
  2. Create a Claude Pro account if you do not have one.
  3. Start a new chat, making sure that you’ve selected Fable as the model. Claude usually defaults to Sonnet.
  4. Upload a copy of your manuscript image and paste the prompt from the kit.
  5. Hit send and you’ll receive your transcription!

2) For purchasing Fable credits

Once you’ve downloaded the toolkit, here’s how to use it if you plan to use the toolkit by purchasing Fable credits:

  1. Go to platform.claude.com.
  2. Create an account by going to Billing and then purchasing pre-paid credits.
  3. Open the Workbench from the Console’s navigation to start a new chat, making sure that you’ve selected Fable as the model. Claude usually defaults to Sonnet.
  4. Upload a copy of your manuscript image and paste the prompt from the kit.
  5. Hit send and you’ll receive your transcription!

Once you receive your transcription, you may notice that agents include the following:

  • [word?] when the agent can see the word but isn’t certain of the reading
  • [b....es] when the agent can make out a partially legible word, where the real letters are visible and the dots represent unknown letters
  • [...] when nothing is traceable at all to the agent

These markers are a design feature to prevent hallucination—it’s better that an agent admits when it doesn’t know something—and are modeled after how human paleographers account for unknowns in their transcriptions. You may see more of these markers if your manuscript page is damaged, if the ink is too faded, or if the image is poor quality.

Perform image-to-text analysis on the manuscript image attached to this message.

Reply with only the text content of the image, following these conventions:
- Keep the original spelling, punctuation, capitalization, and line breaks exactly as written — including u/v and i/j usage, and ff for capital F. Do not modernize anything.
- Silently lower superscript letters (w^th → wth).
- Expand abbreviations, supplying the omitted letters according to how the writer spells the word elsewhere on the page, not modern spelling.
- Where a y-shaped character stands for "th" (as in y^e = the, y^t = that), write th.
- Preserve ampersands (& and &c.) as they appear. Show struck-through words with ~~strikethrough~~.
- Where you cannot read something: use [word?] for an uncertain but visible reading, [b....es] with dots for partly visible letters, and [...] where nothing is traceable. Never fill a gap by guessing from context — an honest gap is better than a plausible word.
- Add no headings, notes, descriptions, or commentary of any kind.

Download