Working With PDFs and Long Documents in Claude
A PDF looks like one simple thing to upload, but two PDFs that look identical on screen can behave very differently once uploaded. One is built from real text and Claude can read every word of it directly. The other is a scan, a photograph of a page saved as a PDF, and what Claude actually receives is closer to an image than to text. Not knowing which kind you're dealing with is the most common reason an answer about a PDF comes back missing something that's clearly on the page.
Tip
If an answer about a PDF seems to be missing content you know is in there, check whether the file is a scan before assuming the question was wrong.
Text PDFs versus scanned PDFs
A text PDF was generated from a document, a Word file exported to PDF, a webpage saved as PDF, a report built in design software. The words themselves are stored as text, and Claude reads them the same way it would read pasted text.
A scanned PDF is an image of a page, produced by a scanner or a phone camera app, even when it looks perfectly clean on screen. Getting text out of an image like that requires a separate step, and the result is sometimes incomplete, especially with small print, handwriting, tables, or a page that was scanned at an angle.
A scanned page and a text page can look identical to you and behave completely differently once uploaded.
A quick way to tell which kind you have before uploading: try selecting a line of text in the PDF viewer. If you can highlight individual words, it's a text PDF. If clicking and dragging just draws a box around the whole page, it's a scan.
A process for a long or uncertain document
- 1
Check whether the content came through completely
After uploading, ask Claude to quote a specific sentence you can see on the page yourself. If it can't find something that's clearly there, the file is likely a scan with incomplete extraction, and it's worth knowing that before relying on any answer from it.
- 2
Reference pages or sections directly
For a document with numbered pages or clear headings, point to them: "on page 4," "in the section on liability." This narrows what Claude focuses on and makes it easier to verify the answer against the source yourself afterward.
- 3
Split a very long document into a sequence of asks
A 60-page report doesn't need one prompt trying to cover the whole thing. Ask about one section, get an answer, ask about the next. Each question builds on the same upload, so there's no need to restate what the document is each time.
- 4
Verify anything you'll rely on against the actual page
For anything that matters, a number, a date, a specific clause, check the answer against the source page directly rather than trusting the extraction was perfect. This takes seconds and catches the rare case where something got garbled or dropped.
Quote the exact sentence on page 2 that describes the notice period for cancellation. If you can't find that section, tell me so directly instead of summarizing the general topic.
”That last instruction matters more than it might look. Asking Claude to say so directly if it can't find something, rather than filling the gap with a plausible-sounding general answer, is what surfaces an extraction problem instead of hiding it.
You know whether the file is a text PDF or a scan
A quoted sentence check confirmed the content came through
Long documents are split into a sequence of section-by-section questions
Anything you'll rely on is checked against the actual page
Inside Claude Tutorial
Verifying against the source applies to every long document, not just PDFs.
A spreadsheet, a long email thread, a set of meeting notes, the same habit of checking a specific claim against the original source before relying on it shows up across the app's lessons, not only with file uploads.
Working across multiple pages of the same document
For a long document, a single question rarely covers everything you need. Instead of trying to design one prompt that reaches every section, treat it as a short back and forth: one section, then the next, each one specific.
Now do the same for the section on data retention, starting around page 15. Same format: quote the exact language, then summarize what it means in plain terms.
”Asking for the same format each time also makes the answers easier to compare once you've gone through several sections, since you're not reconciling three differently structured summaries at the end.
When a scan is unavoidable
Sometimes a scanned document is all you have, an old contract, a handwritten form, a fax someone forwarded years ago. In that case, treat every answer from it as a first pass that needs a manual check, not a final answer. Ask Claude to flag anything it's uncertain about rather than presenting a clean-looking summary that quietly papers over a section it couldn't read clearly.
Go through this scanned form section by section. For any part you can't read with confidence, say so explicitly instead of guessing at what it probably says.
”Continue reading
- Uploading Files Claude Can Actually Use: the general upload habits this article builds on.
- Summarizing a Long Document Without Losing What Matters: once you can reliably work through a long document, shaping what you pull out of it around who's going to read the summary.