Prepare a PDF for ChatGPT or Claude as Markdown

A PDF records where each glyph was drawn, not what it meant, so selecting a few pages in a reader hands you the drawing order rather than the document. Wrapped lines arrive as separate lines, a word hyphenated at the end of a line arrives in two halves, and the running head and page number sit in the middle of your text every page or so.

The converter undoes most of that. Wrapped lines are joined back into paragraphs and end-of-line hyphenation is repaired. Lines set in a larger or bold font become # headings, lines starting with a bullet or a number become list items, and text in a monospace font is wrapped in a fenced code block. The headings are what make a long file usable: they let you send one section instead of the whole document.

The conversion is mechanical. The tool measures font sizes and positions, treats the size carrying most of the text as body text, ranks the larger sizes as heading levels, and looks for bullet characters and columns of short, aligned text. There is no model, no interpretation and no rewriting: every word in the output is a word that was already in your PDF, and the structure around them is inferred from geometry.

The privacy line falls in the same place. Text extraction and structure inference happen inside your browser tab, so the PDF itself is not uploaded. What you paste or send afterward is a different matter: that text leaves your machine and lands with whoever runs the chat service, so check what that service says it does with what you send. If the document is confidential, convert it, cut out only the section you actually need, and send that.

The five checkboxes sit under What to look for. Keep Headings and Lists checked. Keep Tables checked if the document has real tables, but read them afterward: a run of three or more consecutive lines that each break into two or more short, column-aligned pieces is treated as a table, and when the typical piece runs past about forty characters the tool decides it is prose in columns and leaves them as paragraphs.

Leave Drop repeated headers and footers checked. It drops lines in the top or bottom twelfth of the page that repeat on at least sixty percent of the pages, and never on fewer than three, with digits masked so "Page 3 of 40" and "Page 4 of 40" count as the same line. A file shorter than three pages is left alone. Turn the option off only when the running head carries a chapter title you want to keep.

Mark page breaks is off by default and inserts a horizontal rule between pages. Turn it on when you want to quote a page number back later, and leave it off otherwise, since the rules add nothing to the text itself.

The tool takes one PDF at a time and always converts the whole document. For a book or a long manual, send less: use the headings in the preview to find the part you need and copy only that, or split the PDF into chapters first with Split PDF.

Read the preview before you rely on it. Two-column papers are where the output is worst: lines are grouped by their position down the page across the full width, so a line in the left column and the line beside it in the right column are joined into one, and the sentences interleave. Footnotes land in the body near the bottom of a page. No option changes either one, so fix or drop those pages yourself before you send them.

If the tool stops with "No text was found", not one page in the file had any text to read, which usually means a scan. Run OCR PDF on it first: English recognition runs in your browser and lays an invisible text layer over the page images. Its words carry a size but never a bold flag, so headings that were only bold come through as ordinary paragraphs. Recognition is never perfect either, so check any figure or name that matters against the original.