Claude Projects can't read your PDF? The two-minute fix

MakeItMarkdown · July 2026 · 4 min read

You uploaded PDFs to a Claude Project (the same applies to Perplexity Spaces and other AI workspaces), asked a question, and got one of these:

• "I don't see that information in the files" — for a topic that's definitely in there;
• an answer that misquotes a table or mangles a definition;
• several turns of the assistant hunting for which file holds the topic;
• or, with smaller/faster models especially, a flat "I can't read this file."

The instinct is to blame the model. Usually it's the format.

The point: workspaces retrieve chunks, and PDF chunks are glyph soup with no headings — fix the input and the same model finds and quotes correctly.

1 · What's actually happening

A workspace doesn't hand the model your PDF. It extracts text, splits it into chunks, and at question time retrieves the best-matching chunks into the model's context. PDFs sabotage every step: the extracted text has broken reading order and no headings (a PDF stores positioned glyphs, not structure), so chunks split mid-sentence and mid-table, and the search has no section titles to match your question against. The model then answers from two half-thoughts — or declines. Slide-exported PDFs, the classic lecture-notes case, are the worst offenders because their visual layout carries most of the meaning.

2 · The fix

1. Drag your PDFs into MakeItMarkdown — several at once works. Conversion runs in your browser; nothing is uploaded to us.
2. Skim each fidelity report. Warnings tell you if a file has deeper problems (see the caveat below).
3. Download the .zip and upload the .md files to your Project.
4. Remove the original PDFs from the Project. If both versions stay, retrieval sometimes surfaces the bad one.

Then re-ask the exact question that failed. You should see the assistant name the right file immediately and quote it precisely — Markdown headings give the search real handles, and the chunks arrive as coherent topics. We measured this shift in a first-person case study: same workspace, same model, same questions; only the format changed.

3 · The one case this won't fix

If the fidelity report says the PDF looks scanned (image-only, no text layer), there is no text for any converter to extract — that's a different problem with different options, covered in Scanned PDFs: why no text converter can read them. A quick self-test: if you can't select text in the PDF viewer, it's a scan.

4 · Why this also makes the workspace cheaper to run

The Markdown versions are typically a fraction of the size (our measured samples: a 26.6 KB PDF became 0.9 KB), so the storage cap stops being a constraint, and each retrieved chunk carries content instead of layout debris. The full argument, with numbers: Markdown is the best upload format for AI workspaces.