It’s 4:40 on a Friday when a 78-page vendor contract lands in your inbox, with a note asking for a summary by Monday morning. You could read the whole thing. You could also drop the PDF into Claude and ask it to pull out the termination clauses, the payment terms, and anything that looks unusual for a contract of this size. The second option takes about ninety seconds.
That’s the everyday version of Claude AI file analysis. You attach a file, Claude reads it, you ask questions. The interesting part isn’t the upload button — it’s what happens once the document is actually in context, and where the whole approach quietly falls apart if you push it too hard.
What Claude does with a file once you attach it
Claude accepts PDFs, Word documents, plain text, CSV and XLSX spreadsheets, source code files, and images. On the paid plans the context window runs to roughly 200,000 tokens, which is somewhere around 500 pages of dense text. That’s enough to drop in an entire due-diligence folder rather than feeding it one document at a time and hoping it remembers.
It reads pages visually, not just as extracted text
Plenty of document tools strip out the text layer and discard everything else. Claude processes pages as images too, and that difference shows up fast. Scanned contracts, photographed receipts, pages with handwritten margin notes, a slide deck where half the argument lives inside a diagram — all of it stays readable. If you’ve ever pasted a PDF into a tool and gotten back a wall of mangled columns in return, you’ll notice this immediately.
Spreadsheets behave differently
CSV and Excel files work, but they don’t work like a database. Claude is good at explaining what a column is tracking, spotting a broken formula, or summarising trends across a few thousand rows. Ask it to recompute a 40-column financial model cell by cell and you’re asking for trouble — arithmetic drifts on big tables. A better pattern: give it the structure and a sample of rows, ask what questions it would need answered, then do the number-crunching in the spreadsheet itself.
Where file analysis actually earns its keep
The use cases that hold up are the ones where reading is the bottleneck, not thinking.
- Contract and policy review. “List every clause with a liability cap, note the notice periods, and flag anything that differs from our standard template.” Claude returns something you can check line by line in ten minutes instead of an hour.
- Research synthesis. Six papers at once, with a question like “where do these methodologies disagree?” You get a map of the disagreement rather than six summaries you still have to reconcile yourself.
- Messy exports. Support tickets, server logs, a 2,000-row survey dump. Claude can propose categories and cluster the mess faster than you can build a pivot table.
- Code review across files. Attach four source files and it can follow a function call from one module into another, which single-file review misses entirely.
- Screenshot triage. Design feedback, UI bugs, an annotated photo of a prototype. The vision layer does the heavy lifting here.
The prompt decides the quality, not the file
Two people upload the same contract and get wildly different results. One types “summarise this” and receives four paragraphs of pleasant nothing. The other writes: “You’re a commercial contracts lawyer. List every clause that could trigger a payment obligation, quote the sentence, and give the page number. Then tell me what you couldn’t find.”
That second prompt is doing most of the work, which is the core argument in this breakdown of what separates a good Claude AI prompt from a useless one. The habit worth building is asking for structure: a table with fixed columns, a quote alongside every claim, and an explicit list of gaps. That last instruction matters more than people expect — Claude is noticeably more honest about missing information when you invite it to say so.
Limits you’ll bump into eventually
Page numbers are the first casualty. Claude can quote a sentence accurately and still attribute it to page 34 when it’s on page 31. Always ask for the quote itself, then search the document for it.
Long files have a middle problem. Content near the start and end of a very long document gets more reliable attention than the dense section in the middle, so if something crucial lives on page 240 of 300, point Claude at that section directly.
Large tables are the other weak spot. Anything involving arithmetic across hundreds of rows should be verified outside the chat, no exceptions. And if you’re on the free tier, file handling comes with tighter caps than the paid plans — this guide to what Claude AI free actually includes covers exactly where those limits bite.
One non-technical limit worth stating plainly: don’t upload anything you couldn’t forward to a colleague. Contracts with personal data, internal financials, anything under NDA — check your organisation’s settings before the file goes in, not after.
How it compares with ChatGPT’s file handling
Both tools read documents well, and the right pick depends on what you’re doing with them. Broadly, ChatGPT has the edge when you need real computation — its code interpreter will run Python against a spreadsheet and give you back numbers you can trust. Claude has the edge on long narrative documents, visual reading of scanned or diagram-heavy pages, and the tone of its written output. If you’re choosing between the two, this comparison of ChatGPT and Claude gets into the workflow differences rather than the feature lists.
For documents specifically, a useful rule of thumb: hundreds of pages of prose go to Claude, heavy numeric analysis goes to ChatGPT. And if you’re already using ChatGPT and want to sharpen that side of things, there’s a solid walkthrough on getting useful answers from PDFs and spreadsheets with ChatGPT file upload.
A workflow that survives real-world files
Start with an inventory. Before asking for analysis, ask “what’s in this file, section by section?” You’ll catch a truncated PDF or a missing appendix in thirty seconds instead of discovering it three questions later.
Then narrow. Pick one section or one question and go deep. Claude performs better on a focused task with a clear output format than on “review this whole thing.”
Ask for a table, not a paragraph. Clause, page, quote, risk level. Tables are easier to scan, easier to spot gaps in, and much harder for the model to bluff through.
Verify the load-bearing claims. Every number, every quoted sentence, every legal term that changes the meaning of a decision. Treat the first pass as a map of where to look.
When the file is a mess
Real files are rarely clean. A 200-page scan with pages out of order, twelve versions of the same proposal, a screenshot taken at 30% zoom. Three things help here.
Split the job. Ask Claude to inventory the document first and report anything unreadable or duplicated, then hand it sections one at a time. It’s slower per step and faster overall, because you stop chasing hallucinated answers about pages it couldn’t actually read.
Do version comparison in two passes. Single-file mode handles one document well; for “what changed between v4 and v7,” ask for a structured summary of each separately, then paste both summaries back in and ask for the diff. Feeding two long documents in at once works, but the comparison gets thinner as both files grow.
And when a scan is genuinely illegible, say so and stop. Claude will sometimes fill the gap with something plausible, which is the one failure mode that costs real money. A photo taken again in better light beats an hour of prompting around a page the model can’t read.