Stop Feeding AI PDFs - Markdown Is Its Native Format
Idea
AI models read markdown and plain text far more reliably than PDFs, so converting your knowledge into the model’s native format makes its answers more accurate, not just cheaper.
Notes & Verified Quotes
QUOTE: “there are particular types of files that they just find easier to read with markdown being one of them” — src: scratch/sources/claude-skills-side-hustle.txt | source: https://youtu.be/quYKZushRPo t=1214 QUOTE: “they struggle with PDFs more than they do markdown” — src: scratch/sources/claude-skills-side-hustle.txt | source: https://youtu.be/quYKZushRPo t=1224 QUOTE: “You’re making it more accurate and more efficient.” — src: scratch/sources/claude-skills-side-hustle.txt | source: https://youtu.be/quYKZushRPo t=1627 QUOTE: “They just understand markdown better.” — src: scratch/sources/claude-skills-side-hustle.txt | source: https://youtu.be/quYKZushRPo t=1630 QUOTE: “A skill markdown file is just a recipe for an AI agent to follow for a particular task.” — src: scratch/sources/claude-skills-side-hustle.txt | source: https://youtu.be/quYKZushRPo t=1192
- Source: Ryan Dozer on The Kerner Office (TKO Pod). His habit: after a podcast he saves the transcript as a markdown file and feeds that to Claude Code, not a PDF.
- Reason given: models “miss things” reading PDFs; markdown (and HTML) is what they process most cleanly. He couldn’t quote an exact efficiency %, but says a .txt or .md beats a PDF “by a lot.”
- Reframe in the interview: feeding markdown is “translating it to its own language” — accuracy gain, not only a token/cost gain.
- Angle (mine): most people dump a PDF and blame the model when the answer is wrong. The fix is upstream and free — change the format, not the prompt. PDFs are a layout format built for human eyes/printers; the text order, tables and columns get mangled on extraction. Plain markdown is structure the model was trained on.
- Caveat to be honest about: modern models do have decent PDF/vision pipelines, so “never use PDFs” is too strong. The reliable rule: when the content is text you control, give it as markdown/txt; reserve PDFs for when layout itself is the information.
Outline
Draft — LinkedIn (master)
Draft — X (variant)
Final
Related
- Files Are the Universal Interface for Agents — why models are fluent in plain text/markdown but choke on opaque, layout-bound formats
- From Primitives to Domain Tools — agent-native design: meet the model in the format and primitives it actually reasons about
- Your years of expertise are one file away from being a product — sibling from the same talk: a Claude skill is itself just a markdown file