You upload a PDF to your Claude Project and ask questions about it. Instead of clear answers, you get vague statements, wrong quotes, or an open "That is not in the documents." The problem is rarely Claude. It is usually how the PDF was loaded into the project. This guide shows you the four most common mistakes and how to fix them concretely.
Mistake one: uploading scanned PDFs without OCR
This is by far the most common mistake. You scan a document, save it as PDF, upload it to Claude Project. But Claude does not get text, only images of text. The result: Claude acts as if there is no information.
How to recognize whether your PDF is scanned:
Open the PDF in a reader and try to highlight text. If you can mark and copy words, the PDF is text-based. If only a rectangle is highlighted or you cannot select anything at all, the PDF is a scanned image.
Fix:
Run the PDF through OCR before upload. You have three options:
- MarkItDown in the browser automatically detects scanned PDFs and runs OCR without any configuration on your part.
- Adobe Acrobat has a "Recognize Text" command that performs OCR locally.
- Tesseract as a command-line tool for power users.
After OCR treatment, the PDF contains real text that Claude can read and understand.
Even with good OCR quality, it is better to convert the PDF directly to Markdown instead of uploading it as a PDF. Markdown gives Claude a clearer structure.
Mistake two: large PDFs with unstructured content
You upload a 200-page PDF to the project. Claude can theoretically read all 200 pages, but for specific questions he often cannot find the relevant section or mixes content from different sections.
The problem: Claude performs internal retrieval for large knowledge bases. If the PDF has no clear structure, no table of contents, no clear sections, no sensible hierarchy, this retrieval fails.
Fix:
Break large PDFs into thematically clearly separated Markdown files before uploading. Instead of one file company-handbook-complete.pdf, ten files:
01-company-history.md
02-organizational-structure.md
03-processes-sales.md
04-processes-it.md
05-policies-employment.md
06-policies-data-protection.md
07-tools-software.md
08-onboarding-new-employees.md
09-meeting-formats.md
10-contacts-stakeholders.md
Each file has a clearly defined topic area. When someone asks "How does onboarding work?", Claude immediately knows which file to look in.
A rough rule of thumb: no knowledge file should be longer than 30 Markdown pages. If it is, split it further.
Mistake three: multi-column PDFs without preprocessing
Scientific papers, magazines, annual reports, and many corporate documents have two- or three-column layouts. When Claude extracts these PDFs as text, he often reads in the wrong order. The top of the left column gets mixed with the top of the right column, instead of reading the left column completely from top to bottom.
The result: disconnected text fragments that Claude processes semantically incorrectly.
Fix:
Convert the PDF before upload with a tool that handles multi-column layouts. MarkItDown attempts to reconstruct the correct reading order. For very complex layouts (magazines with boxes, sidebars, embedded diagrams) there is no perfect tool. Here only helps:
- Convert PDF to Markdown
- Manually go through the resulting Markdown
- Fix wrong order
- Only then upload
For ten such documents, that is one hour of work. For a hundred such documents, you should invest in a dedicated PDF processing tool like PyMuPDF.
If you get answers from Claude that jump between topics mid-sentence, that is a strong indicator of incorrectly read multi-column PDFs. In that case, check the reading order in the converted text.
Mistake four: no metadata at the document start
You upload a PDF with the filename report.pdf. Inside is a report, but Claude does not know what about. The title on the first page is in small print, the context (who wrote it, when, for whom?) is missing.
For a single PDF, this is not a big problem. With twenty PDFs in the project, it becomes chaotic. Claude cannot reliably tell which document belongs to which topic.
Fix:
When you convert documents to Markdown (which you should anyway), add a metadata block at the start of each document:
# Quarterly Report Q4 2025
**Created by:** Anna Müller, Head of Sales
**Date:** January 15, 2026
**Content:** Revenue development, customer acquisition, top products, Q1 outlook
**Audience:** Management, investors
---
[Content follows here...]
These five lines at the beginning make a big difference. When Claude gets a question about Q4 revenue, he immediately finds the right file. When someone asks "Who wrote the last quarterly report?", Claude can answer instead of saying "That is not in the documents."
Bonus: what you should never do
Three things that should never be uploaded to Claude Projects:
One, password-protected PDFs. Claude cannot decrypt them. Remove the password first.
Two, PDFs with embedded forms. The form data often gets lost, the remaining text gets distorted by the form structure. Convert to Markdown.
Three, PDFs from PowerPoint exports without content substance. An 80-page PDF from a presentation with bullet points that only say "leverage synergies" and "best-of-breed solution" is worthless to Claude. No converter helps here, only the original speaker script of the talk.
A concrete workflow for practice
If you regularly use PDFs in Claude Projects, you should establish a fixed workflow:
- Receive or create PDF
- Check whether OCR is needed (test text selection)
- Convert PDF on MarkItDown to Markdown
- Open converted Markdown and check briefly (reading order, tables, lists)
- Add metadata block at the start
- If document is large, split into thematic parts
- Name files with prefix numbers (01-, 02-, 03-)
- Upload to Claude Project
- Test with three typical questions whether retrieval works
The entire workflow takes five to fifteen minutes per document and significantly improves Claude's answer quality.
Conclusion
Most problems with PDFs in Claude Projects are not Claude problems but preparation problems. Four main mistakes cover 90 percent of all cases: no OCR on scans, large unstructured documents, multi-column layouts without preprocessing, missing metadata. If you consistently avoid these four points, you get noticeably more precise answers in every project without upgrading to a larger Claude plan.
Read more
- Convert PDF to Markdown for preprocessing
- Feeding Claude Projects with documents for the comprehensive guide
- What is Markdown?, if you haven't read the basics guide yet
- Use MarkItDown directly for browser conversion