Files and documents

Files and documents are the PDFs, spreadsheets and images you attach to a chat for Claude to read, and it reads some of them far better than others. Attaching a document is the point at which Claude stops being a clever text box and starts being useful for your actual job. It is also where the most confident wrong answers come from, because there is a gap between what you can see in a file and what it can see.

This lesson is mostly that gap.

What can you attach?

Documents in the usual formats: PDF, Word, plain text, HTML, CSV, JSON, EPUB, ODT, RTF. Images as JPEG, PNG, GIF or WebP.

Excel files need one setting turned on first. It is called Code execution and file creation, it lives under Settings > Capabilities, and it is on every plan including free. On a work account it is Organization settings > Capabilities and an administrator may have it switched off. Turn it on now; the spreadsheet section below needs it.

The limits, as documented:

• 500 MB per file, and up to 20 files in one chat.
• Images up to 8000 by 8000 pixels.
• PDFs up to 1000 pages. Over that it refuses.
• Files added to a project’s knowledge are capped lower, at 30 MB each.

You will not meet most of those. The one worth remembering is the page count, because of what happens either side of it.

The three things it cannot see

Here is the part that matters, and none of it is announced.

In a PDF over 100 pages, it stops looking at the pages. Up to 100 pages it reads both the text and the visual layout. From 101 to 1000, it processes the text only. So a chart on page 300 of a long report is invisible, and the answer you get about it will be assembled from the surrounding words.

In a Word document, it cannot see the pictures at all. For anything that is not a PDF, the documentation is blunt: Claude extracts text only, and “if they contain embedded images, Claude won’t be able to read or interpret them”. An embedded chart, a screenshot, a scanned signature: it does not know they are there.

In a scanned document, there may be no text at all. A PDF made by a scanner is a photograph of a page. Under 100 pages it can look at it as an image and often does well. Over 100, there is no text to extract and nothing to look at, and you will get an answer built from very little.

The general rule that covers all three: if the information you want is a picture rather than words, ask yourself whether it can see pictures in this particular file. The answer depends on the format and the length, and it is never mentioned in the reply.

Spreadsheets are a different job

A spreadsheet is not a document, and treating it like one is the most common mistake in this lesson.

For a small sheet, a few hundred rows, attaching it works and you can ask ordinary questions about it.

For anything larger, or anything involving arithmetic across many rows, you want Code execution and file creation switched on, because then it writes a few lines of code and runs them rather than reading numbers off a page. The difference is the difference between working out a total and estimating one. It is also what lets Claude hand you back a real .xlsx, .docx or .pdf rather than text you have to paste somewhere.

The tell that you are on the wrong side of this: you asked for a sum and got a number with no working. Ask “how did you get that” and see whether the answer is a calculation or a paraphrase.

Spot-checking, which takes fifteen seconds

Whenever an answer depends on a document, check one fact against the file before you use any of it.

Pick a number or a name from the answer, find it in the document yourself, and see whether it is where the answer says it is. Not to be suspicious; to find out whether it read what you think it read. If the number is right, the extraction worked. If the number is right but the page reference is wrong, it is working from text and cannot see the layout, which tells you exactly which of the three problems above you have.

The habit is: check one fact, then trust the rest. Checking every fact is slower than doing the work yourself, and checking none is how people end up quoting a figure that was never in the document.

Your turn

You are going to make the Word-document blind spot happen on purpose, because it is the one that catches people who have no reason to suspect it.

Take any document with a chart or a picture in it. If you do not have one, make one: a Word document with two paragraphs and a screenshot of anything pasted in the middle.

1. Attach the Word version and ask:

   Describe everything in this document, including anything that is not text.

2. Now save the same document as a PDF and attach that in a new chat. Ask the identical question.

Check: compare the two answers. The documented behaviour is that the PDF answer describes the picture and the Word answer does not, without telling you there was a picture it could not see. Same content, same question, different answer, and nothing on screen explaining why.

If your Word answer does describe the picture, that is worth knowing too: the extraction has improved since this was written, and the blind spot has moved rather than gone. Try the same test with a scanned PDF of more than a hundred pages, which is the version of it that is hardest to fix.

Then, on any document you have handy, do the fifteen-second spot check:

Summarise this and give me the three most important numbers in it, and say where each
one appears.

Check: find one of those three numbers in the document yourself. It is where the answer says it is, or the answer is working from text without layout. Either result tells you something you can use.

Recap

Most formats attach and work. The limits worth knowing are 20 files per chat and 1000 pages per PDF.

Three blind spots, none of them announced: a PDF over 100 pages is read as text only, a Word document has its pictures dropped entirely, and a scan may contain no text at all. If what you want is a picture rather than words, work out whether it can see pictures in this file.

Spreadsheets need the code setting for anything that involves arithmetic across rows. A sum with no working is a sum you should ask about.

Check one fact against the document before you use the rest. Not every fact, and not none.

Next: when Claude makes the thing instead of describing it.