Working with Images in Claude Code

The issue: Most guidance on Claude for research focuses on text files — PDFs, DOCX, markdown notes. But historians and archivists work with visual material: manuscript photographs, historical maps, archival diagrams, tables in scanned sources, infographics in papers. Claude can read images directly — but the capabilities and limits are different from text, and the workflows require adjustment.

What Claude can do with images

Describe and interpret visual content. Show Claude a historical map and ask what it depicts, what region, what period indicators are visible, what is unusual. Show it a network diagram and ask it to describe the structure. Show it a page from an illuminated manuscript and ask what scene is depicted. Claude has broad visual knowledge and can engage with content at this level reliably.

Transcribe readable handwritten text. For legible hands — clear Carolingian minuscule, secretary hand, well-preserved Gothic bookhand — Claude can transcribe with reasonable accuracy. It is not a specialist OCR tool, but for a single page of moderately clear script, it can produce a working transcription you then verify and correct.

Extract structured data from visual tables. A table photographed in a manuscript, a chart in a scanned report, a data grid in a historical document — if the table is readable, Claude can extract the contents as structured text (CSV, markdown table) you can then work with. This is often faster than manual transcription for tabular material.

Identify visual elements and compare images. Ask it to identify the type of seal on a document, describe the iconography of a miniature, or note the differences between two manuscript pages. It can describe, classify, and compare within the limits of its visual training.

Answer questions about image content. "What names are visible in the margin of this page?" "What does the key of this map indicate?" "Are there any annotations in a different hand?" Claude reads the image and answers — within the limits described below.

What Claude does less well

Heavily degraded or damaged text. Faded ink, water damage, overwritten text, tight binding — when the script is unclear, Claude's transcription degrades. It will still attempt a reading, but the confidence is lower and errors are harder to catch because the plausible-sounding output looks like it might be right.

Non-Latin scripts without specialist training. Hebrew, Arabic, Greek, Cyrillic, and other scripts Claude can handle to varying degrees, but accuracy on historical manuscript hands in these scripts is less reliable than for Latin-alphabet hands. Treat transcriptions as a starting point requiring specialist review.

Very long documents as single images. A photograph of a full double-page spread with dense text will produce more errors than the same content in two separate close-up images. Closer crops of specific sections get better results.

Searching within image content. Claude reads the image you give it; it cannot search an image the way it searches text. If you have 40 manuscript photographs and want to find all that mention a specific name, the workflow is: transcribe them all first, then search the transcriptions. The image-reading step is upstream of the search step, not a replacement for it.

Precise measurements or spatial coordinates. Claude can describe layout and relative positions, but it cannot give you precise pixel coordinates, map projections, or geometric measurements. For spatial analysis of historical maps, specialist GIS tools are still needed after Claude's descriptive pass.

Practical workflows

Single-image analysis

Copy the image file into your project folder. Reference it by filename:

Read 01_deposition_bernardus.jpg

1. Is the hand legible? Rate it: clear / moderate / difficult.
2. Transcribe the first 5 lines.
3. Are there any marginalia visible?
4. What writing support does this appear to be (parchment, paper)?

Claude will read the image and respond. Check the transcription against the original before relying on it.

Batch transcription of a folder of scans

I have 15 manuscript photographs in /scans/
They are all depositions from the same inquisition register, c. 1240s, southern France.
The hand is consistent throughout.

For each image:
1. Assess legibility (clear / moderate / difficult)
2. Transcribe the first paragraph (approximately the first 8 lines)
3. Note any visible proper names

Write results to transcription_batch.md — one section per file.

Review the output. For the "moderate" and "difficult" files, do a second pass with closer crops if available.

Extracting data from a visual table

Read this scanned page: tax_register_1342_folio_12r.jpg

There is a table on this page listing taxpayers and their assessments.
Extract the contents as a markdown table with columns:
Name | Amount | Notes

If any cells are unclear, mark them [?].

Comparing two versions of the same source

I have two photographs of the same document, taken at different times:

- ms_x_1970.jpg (older photograph, black and white)

- ms_x_2024.jpg (recent photograph, colour)

Compare them. What differences do you notice?
Are there visible losses, additions, or changes in condition between the two photographs?

DISSINET-specific applications

Network source documents. Inquisition depositions, notarial records, court registers — when you have a photograph rather than a digital transcription, Claude can do an initial reading pass, identify personal names, and flag uncertain readings before you transcribe formally.

Map analysis. Historical maps of the Languedoc, Cathar settlement maps, diocesan boundary maps — Claude can describe, identify features, and compare with modern geography. It will not georeference or produce GIS-compatible output, but it can give a descriptive reading that orients your analysis.

Visual cross-referencing. If you have both a manuscript image and a transcription, show Claude both and ask it to verify the transcription: "Compare this photograph with this transcribed text. Note any discrepancies." This is a useful first-pass error-check before formal palaeographic review.

Key limitations to keep in mind

Image quality sets the ceiling. Claude cannot read what is not visible. Poor lighting, low resolution, and physical damage produce proportionally worse results. If you have access to better photographs (UV, multispectral, high-resolution), they will produce better Claude output.

Transcription is a starting point, not a product. For scholarly use, Claude's image-based transcription requires human review. It is useful for: initial orientation, identifying what a document contains before deciding whether to transcribe it fully, first-draft transcription that you correct rather than produce from scratch. It is not a replacement for trained palaeographic reading.

Claude cannot cite its image readings. Unlike its text work, where you can verify what it said against the source, image readings produce text that is harder to audit. Build in a verification step before relying on any image-derived data in analysis.

Related