Claude Multimodal

What Claude Can See

Claude's vision capabilities allow it to understand images at a semantic level — not just identify what's in them, but understand the meaning and context.

Documents and forms: Claude can read printed and handwritten text in images, understand document layout, and extract specific fields from forms, invoices, receipts, and certificates. This is the foundation of AI document processing.

Charts and data visualisations: Claude can read charts, graphs, and tables — understanding the data they represent, identifying trends, and comparing values. Useful for analysing screenshots of dashboards, reports, and presentations.

Product images: Claude can describe products, identify features, compare variants, and detect defects — enabling visual product cataloguing, quality inspection, and e-commerce applications.

Screenshots and UIs: Claude can read and interact with screenshots of applications, websites, and interfaces — useful for automated testing, UI documentation, and user experience analysis.

Handwritten content: Claude reads handwritten text with reasonable accuracy, enabling processing of handwritten forms, notes, and signatures.

Document Processing with Vision

The most commercially valuable vision use case is document processing. Physical documents — paper invoices, scanned contracts, photographed forms — have historically required expensive OCR software and complex template matching. Claude's vision capability handles these naturally.

Invoice processing: Photograph or scan an invoice, send the image to Claude with an extraction prompt. Claude extracts: vendor name, invoice number, date, line items, amounts, and payment terms. The structured data is returned as JSON and written to your accounting system.

For an accounts payable team processing 500 invoices per month, this automation saves 40-60 hours of manual data entry per month.

Form processing: Any form that arrives as a physical document — application forms, customer intake forms, government forms — can be processed by Claude's vision API. The same prompt-based extraction approach works for any document type.

Receipt processing: Expense management workflows use Claude to extract data from receipt photos — vendor, amount, date, category. Submit a photo of a receipt; the expense system is updated automatically.

Implementation

To send an image to Claude via the API, include the image in the content array of the user message:

Supported image formats: JPEG, PNG, GIF, WebP. Maximum image size: 20MB. Images can also be passed as URLs if the API can access them.

Vision Use Cases in Automation

In n8n and Make.com, vision automation workflows follow the same pattern as text workflows, with the additional step of handling the image file:

1. Receive image via email, upload form, or file watch
2. Convert to base64 (or get URL if stored in S3/Drive)
3. Call Claude API with image + extraction prompt
4. Parse the JSON response
5. Write structured data to destination system
6. Route low-confidence extractions to human review

Limitations

No handwritten accuracy guarantee: Claude reads handwritten text reasonably well, but accuracy varies with legibility. For handwritten documents, build in human review for critical fields.

Image quality matters: Blurry, dark, or low-resolution images produce worse extraction quality. Establish minimum image quality requirements for automated processing.

Not real-time video: Claude processes static images, not video frames or live video. For video analysis, extract frames first.

No rendering: Claude can't render PDFs or complex documents natively. Convert to images (screenshots or PDF-to-image conversion) before sending.

Recommended Tools

• Anthropic API — Claude vision API
• n8n — Image-based document processing workflows
• Make.com — Visual automation for image processing
• AWS S3 — Image storage for high-volume processing
• Airtable — Review queue for low-confidence extractions

“Claude's vision capability turns every document your business handles into machine-readable structured data — automatically.”