Troubleshoot Claude when it cannot see a file or finish a task
Claude summarises the 140-page tender response in seconds and says it contains no unit rates. They're on page 112, in a table. Before you retry the whole job, find out what it could actually see.
What you'll make
a one-line diagnosis, one checked sample from the file, and a next step that works from what Claude could really read.
What you'll need
Claude in your work account, the file or task that went wrong, the original to compare against, and about fifteen minutes.
Example result.
Observed: summary of a 140-page PDF said "no unit rates provided"; rates are a table image on p.112.
Likely cause: PDFs over 100 pages are read as text only, so images and charts are skipped.
Test: asked for the rates on p.112 — Claude said it couldn't read that page's table.
Next step: split pages 101–140 into a separate PDF; rates read and checked against the original.1. Write down what you can see
Record the task, the file it needed, what you expected and what you got. Note any error message, and whether Claude is still working, waiting for you or showing a limit. Save useful output before restarting: repeating a large request uses more of your limit without fixing an unreadable file.
Why this works: “it can’t see my file” covers a dozen different causes, and the symptom narrows them.
What you should see: one sentence describing what happened, not why.
2. Match the symptom to its usual cause
| Symptom | Usual cause | First check |
|---|---|---|
| Misses tables, charts or images in a long PDF | PDFs of 101–1,000 pages are read as text only | Split it into parts of 100 pages or fewer |
| Can’t read a Word or text file’s pictures | Non-PDF documents are text only | Export the page as PDF or an image |
| “Uploaded file is too large” | Over 1,000 pages, or over the file size limit | Split the file |
| No downloadable file, just a description | Code execution and file creation is off | Settings › Capabilities (on Team or Enterprise, ask your owner) |
| Stops mid-task with a limit message | Usage limit reached; long chats use more | Wait for the reset, or start a new chat with a short handover |
| Can’t reach Drive or another app | Connector not connected | + › Connectors › Manage connectors |
| A cloud task can’t read your laptop folder | Claude reaches connected folders only through the desktop app, while it’s open | Reopen the desktop app |
Why this works: most “can’t see” problems are a documented limit with a one-line check.
What you should see: one likely cause and one check to run.
3. Ask for evidence of what it read
The problem is [what happened]. This task needs [file and version]. Before continuing, tell me which files and pages you can access, and show [one short passage or a few values] from [specific page or section]. Say plainly what you couldn't read, including images or tables. Do not invent a sample or fill gaps from context.[what happened] comes from step 1; the other brackets name the file, a value you can check and where it is. Pick the page where you know the answer.
Why this works: a correct value from the right page is evidence; “I’ve reviewed the whole document” isn’t. One correct sample proves that sample, not the whole file.
What you should see: a value you can compare with the original, or a clear statement that it couldn’t read that part.
4. Continue from checked material
Fix the cause from step 2 (split the PDF, switch on file creation, connect the app) and repeat the step 3 test. If you can’t fix it, give Claude the part it can read and make the gap visible. If it’s still stuck, check Claude’s status page, then report to your admin with the account type, surface and app version, date and time, the exact error and a small example.
Why this works: re-running the same small test after each change tells you whether that change fixed it.
What you should see: the step 3 sample now correct, or a short report ready to send.
Check the result
- The answer quotes something from the part you were worried about.
- Anything Claude couldn’t read is listed, not silently skipped.
- You changed one thing at a time.
- No unrelated folder, account or app was connected to get round the problem.
The usual failure: a confident absence. “The document doesn’t include unit rates” sounds like a finding, but it can mean “I couldn’t see that page”. Spot it by testing the one place you know the answer is. Then continue only from what’s checked:
Use only [checked file or excerpt]. Produce [small result] and keep [fixed facts] as they are. Mark anything you couldn't read as unresolved. Do not say the information is absent if you couldn't read that part of the original.Name the checked material, the result you want and the facts that must not change.
Worked example: the unit rates on page 112
Fictional example. Nadia, a procurement analyst, uploads Harlow Facilities’ 140-page tender response and asks for a summary of prices. The unit rates are a table pasted into page 112 as an image.
Claude’s summary said: “Harlow provides a fixed annual fee of £84,000. No unit rates are provided.”
- The failure. A confident absence. The fixed fee is on page 9, in text; the rates are an image on page 112. At 140 pages, the PDF is read as text only, so the image was skipped.
- The diagnosis. The step 3 test asked for the rates on page 112. Claude said it couldn’t read the table on that page, which matches the first row of the table.
- The repair. Nadia saved pages 101–140 as a separate PDF and uploaded it. The same test now returned the cleaning rate as £18.40 per hour, which matched the original. The revised summary gave both the annual fee and the unit rates, and listed nothing as unread.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Cloud tasks and your computer’s folders
Claude Cowork on Pro and Max plans now runs new tasks in the cloud, so work continues after you close your laptop. Your folders stay on your computer: Claude reaches only folders you’ve connected, through the desktop app, and only while it’s open. A task that keeps running isn’t proof it can still read your files. Give Claude a folder for one bounded job covers setting this up.
Long chats and usage
Usage limits reset every five hours with a weekly limit on top, and everything (web, desktop and Claude Code) counts towards the same limit. Long chats that trigger automatic context management use more of it. If you’re close to the limit, start a new chat with a short summary rather than retrying in the long one. For the same checks in ChatGPT, see Troubleshoot ChatGPT when a file, feature or response is missing.