Why Claude rewards Markdown more than other models
Claude is unusually good at long, structured documents — but only when the structure is real. Anthropic's training pipeline leans heavily on documents that already use Markdown headings, so Claude treats ## like a section break, > like a quote, and a fenced code block like inviolable code. Feed it PDF's reflowed text and you get the opposite: Claude tries to infer structure from indentation and spacing, and gets it wrong on the long tail.
The practical impact: on Sonnet 4.6 and Opus 4.7, Markdown input produces noticeably more confident citations ("from Section 4.2…"), better quote extraction, and fewer "I cannot find this in the document" responses on questions whose answers are obviously present.
Workflow with Claude Projects
Claude Projects is the killer feature for repeated work on the same document set. Convert each PDF once, drop the Markdown into the Project's knowledge base, and every Claude.ai conversation in that Project starts with clean structured context. Re-converting only when the source PDF changes keeps your cost essentially zero.
For one-off questions, paste the Markdown into a single message instead of attaching the PDF — Claude's file parser is good but it still costs tokens you don't need to spend.