Claude handles PDFs well — but it handles Markdown better. Convert your PDF first and Claude can cite sections by heading, follow lists exactly, and reason over the document without re-deriving its shape from raw text.
Claude is unusually good at long-context tasks: 200K-token windows, faithful citations, careful reasoning. The bottleneck on a PDF-driven conversation isn't usually the model — it's the input. PDFs hand Claude a flat stream of words and force it to infer the structure. Markdown hands Claude the structure directly: H2 means section, H3 means subsection, a bullet list means peer items. The result is fewer hallucinated section names, more faithful citations, and shorter answers (because Claude isn't spending output tokens explaining what it inferred about the shape).
For the claude.ai web or desktop app: drop your PDF on t0md, copy the Markdown, paste into a new conversation. Add a one-line system prompt like "This is the source document; cite by heading when answering" and Claude will reference sections back at you exactly. For Claude Code (the CLI), use the MCP server instead — see the Claude Code guide.
Two reasons. First, you can see exactly what Claude is reading and fix any extraction issues before they affect the answers. Second, Markdown is more compact than the underlying PDF text Claude extracts, leaving more room in the context window for reasoning.
No — Markdown has no page concept. Citations will point to headings, which is usually more useful anyway: "per the Methodology section" beats "per page 14" in any document longer than a few pages.
Yes. Upload the .md file (or paste the Markdown) into the project's knowledge. Claude will index it the same way as any other text source.