Markdown for LLMs: How to Prepare Documents for ChatGPT, Claude and RAG

Convert documents to Markdown before giving them to an AI model. Markdown keeps the structure that matters — headings, lists, tables, links — while dropping the layout code that wastes tokens, so ChatGPT, Claude or a RAG pipeline reads the content accurately and more of it fits in the context window. Use a converter for each source type (PDF, Word, PowerPoint, web pages), check the output, then split long documents at their headings.

1. Structure survives. A heading is ##, a list item is -, a table is rows of pipes. The model sees which text is a title, which is a step and which numbers belong to which column — information that is often lost when a PDF is flattened to raw text.

2. Fewer tokens. Models read and bill by token (roughly four characters of English text per token). HTML spends most of its characters on tags and attributes; Markdown spends almost none. We ran a typical pricing block — a heading, a sentence, a three-row table and two links — through our HTML to Markdown converter: 575 characters of HTML became 227 characters of Markdown, about 60% smaller, with nothing meaningful lost.

3. You can see exactly what the model sees. A PDF can look perfect on screen while its text layer is out of order. Converting first lets you spot scrambled columns, missing tables or repeated page headers before they confuse the answer.

4. Models are fluent in it. Much of the well-structured text models learn from — documentation, READMEs, Q&A sites — is Markdown, which is also why they answer in it.

Every converter on MDConverter runs in your browser, so a confidential document is not uploaded anywhere while you prepare it — you decide what is sent to the AI.

1. Remove page furniture. Repeated headers, footers, page numbers and "Confidential — do not distribute" lines add noise to every chunk.
2. Fix heading levels. One # title, then ## sections and ### sub-sections. Models and chunkers use these as boundaries.
3. Keep tables as tables. If a table came out as loose lines of numbers, rebuild it — a model cannot reliably tell which value belongs to which column otherwise.
4. Describe images that matter. Converters keep image alt text, not the picture. Add a sentence for any chart or diagram the answer depends on.
5. Delete boilerplate. Legal disclaimers, tables of contents and cover pages rarely help and cost tokens.
6. Check special characters. Ligatures (fi), hyphenation at line ends and smart quotes from PDFs can break search and matching.

For a knowledge base or a retrieval pipeline, Markdown headings give you natural chunk boundaries:

• Split at ## or ### headings first, then only split further if a section is still too long. A few hundred to about a thousand tokens per chunk is a common starting point; tune it against real questions.
• Prefix each chunk with its heading path (for example Employee handbook > Leave > Parental leave) so a chunk still makes sense when it is retrieved on its own.
• Never split inside a table or a code block. Keep them whole, even if that chunk runs long.
• Keep links. A cited URL in the chunk lets the model point readers to the source.

If you need batch processing, OCR and layout analysis across thousands of files, an open-source pipeline such as Docling is the better fit; browser converters are best for individual documents you want to check by eye.

Markdown also makes prompts clearer. Headings and fenced blocks separate your instructions from the material, so the model does not confuse the two.