A PDF that changes based on who is reading

The article describes a technique for creating PDFs that render identically to human readers but provide clean, structured markdown to machine readers like LLMs, using a PDF specification property for replacement text that has been largely unused since 2001. By attaching marked-content sequences to the PDF's content stream, extractors receive properly formatted markdown with headings, tables, and lists instead of jumbled text with broken line wraps and flattened structure. While token counts remain similar between normal and adaptive PDFs, the structured output eliminates the need for LLMs to guess document hierarchy and relationships.