Claude: Zero to Hero
Foundations · module 4 of 8

Files, images, and vision

You are here · All paths · Free plan (some limits) · 40–50 min · Assumes Prompting fundamentals.

What you’ll learn: how to get material into Claude effectively, and how to work with images and PDFs without wasting context or getting hallucinated readings.

If you only read one thing

You can hand Claude PDFs, images, screenshots, spreadsheets and code, by dragging, pasting, or using the attachment button. It reads them properly — including the visual layout of a PDF, not just the text.

Three things will save you most of the trouble people have here:

What you can upload

In the chat apps you can attach:

You can drag and drop, paste directly (screenshots especially), or use the attachment button.

Limits on file size and count vary by plan and change over time — see Upload files to Claude.

How Claude actually reads a file

This is worth understanding because it explains several confusing behaviours.

Text files are converted to text and placed in the context window. Straightforward.

PDFs are handled two ways depending on the surface and the document. Text-based PDFs get extracted as text. PDFs with meaningful visual layout — charts, scanned pages, complex tables — are also processed as images, so Claude can see the layout. Image processing costs substantially more tokens per page.

Images are converted into tokens based on dimensions. A large screenshot can cost a few thousand tokens. Resizing a 4000px screenshot down to 1500px before uploading saves real money and usually loses nothing.

Spreadsheets are converted to a text representation. For anything with formulas, pivot logic, or many sheets, this loses fidelity — use Claude for Excel (see Claude inside your apps) or ask Claude to write code against the file instead.

The consequence

Everything you attach consumes context on every subsequent turn. Attaching a 300-page PDF and then having a 30-turn conversation is expensive and, past a point, degrades quality.

Better patterns:

Working with documents well

Ask for evidence, not just answers

The quote-grounding pattern from Prompting fundamentals matters most here:

Find the passages in the attached contract relevant to termination rights. Quote them verbatim in <quotes> tags with page numbers. Then, based only on those quotes, answer: under what circumstances can either party terminate early?

This turns “Claude summarised the contract” into “Claude located and cited the clauses”, which is a different reliability class entirely.

Multiple documents: label them

<documents>
  <document index="1">
    <source>2025-contract-v3.pdf</source>
    <document_content></document_content>
  </document>
  <document index="2">
    <source>2024-contract-signed.pdf</source>
    <document_content></document_content>
  </document>
</documents>

What changed between these two contracts? Cite the document index for each difference.

Without labels, Claude will conflate them. With labels, it can cite.

Verify anything that matters

Claude is very good at reading documents and still occasionally wrong. For anything consequential — legal, financial, medical, compliance — treat its output as a first pass by a capable assistant, not as ground truth. Ask for citations and check them.

Vision

Claude can see images. Practically, that means:

Use case Works well
Reading text in a screenshot Yes — including handwriting, usually
Describing a photo Yes
Reading a chart or graph Yes, with care — verify the numbers
Debugging a UI from a screenshot Yes, this is genuinely useful
Reading a diagram or architecture drawing Yes
Extracting a table from an image Yes
Precise pixel measurements No
Counting many small objects Unreliable
Distinguishing very similar colours Unreliable

The crop technique

Anthropic’s own testing shows a consistent uplift when Claude can “zoom in” on relevant regions of an image. If you’re building something image-heavy, give Claude a crop tool. In chat, do it manually: crop to the region you care about before uploading. A cropped 800px region beats a full 4000px screenshot both for accuracy and for cost.

Multiple images

Current Opus models handle multi-image contexts notably better than earlier generations. When you upload several, label them:

Image 1 is the current design, image 2 is the proposed redesign. List every difference in the navigation.

Video

Not directly supported — but you can split a video into frames and upload those. This works surprisingly well for short clips.

Try it

Exercise 1 — Grounded document QA. Take a real PDF from your work (10+ pages). Ask a factual question directly. Then ask with the quote-extraction pattern. Verify both answers against the document. Note whether the ungrounded one was right, and whether you could have told it was right.

Exercise 2 — Screenshot debugging. Screenshot any website with a layout you find ugly or broken. Ask Claude what’s wrong and how to fix it. Then ask for the CSS.

Exercise 3 — Chart extraction. Find a chart image (no underlying data). Ask Claude to extract the data as a table. Check it against the axes yourself. This calibrates how much you should trust chart reading.

Exercise 4 — Crop test. Take a dense screenshot — a full spreadsheet or dashboard. Ask a question about one small region. Then crop to that region and ask again. Compare accuracy.

Exercise 5 — Context cost. Attach a large PDF and have a 15-turn conversation. Then start fresh, extract the three relevant pages, and have the same conversation. Note the speed difference.

Checkpoint

Going deeper