.rec is a new open format that converts PDFs into something AI can actually consume efficiently.
~100 tokens per page instead of ~1,600, with page references, embedded metadata, and a verifiable schema —
structure that raw Markdown simply doesn't have.
Same document. Same information. 16× cheaper.
Drop any PDF below — get a .rec file ready to paste into Claude, ChatGPT, or Gemini.
Drop your PDF here
or
PDF only · any size
Converting…
Conversion complete
Who is this for?
You feed papers, theses, and reports to AI for literature reviews, summaries, or Q&A. A single 40-page paper costs ~64,000 tokens as PDF. With .rec, that's under 4,000. Process entire libraries without burning through API credits.
You use AI to understand lecture notes, textbooks, and course materials. Free-tier context windows fill up fast. .rec gets 16× more content into the same window — so your whole semester's notes fit in a single prompt.
On ChatGPT, Claude, or Gemini free plans, every token matters. You hit limits before you finish a document. .rec shrinks the document 16× so you can ask more, go further, and stop getting cut off mid-analysis.
pdfplumber extracts headings, paragraphs, tables, lists — preserving structure but discarding layout, fonts, and images.
Each page is serialized to compact JSON, zlib-compressed, and packed into a binary envelope with a SHA-256 checksum.
Pass the .rec file to Claude, GPT-4o, or Gemini via the API helper. The model reads it as structured text at a fraction of the token cost.
Two steps. That's it.
This plain-text file tells any AI how to read .rec documents. Paste its contents into your AI's system prompt — once, permanently. Works with Claude, ChatGPT, Gemini, or any LLM.
Installs the rec-convert command on your machine.
After this, convert any PDF to .rec from the terminal.
Same information, a fraction of the tokens. Every page, every time.
Why not just Markdown?
Token cost inside the context window is nearly the same. The difference is in the transport layer — structure, provenance, and verifiability that Markdown simply doesn't have.
Every block is tied to its source page ([P9]). Citability is structural — not an afterthought. Markdown has no concept of pages.
Title, source file, language, creation date, and page count travel inside the file as standard fields — not a frontmatter convention.
SHA-256 checksum lets you verify the document arrived complete and unmodified. Markdown has no such mechanism.
Every block is typed JSON. {"t":"fn"} is always a footnote, everywhere. No dialects, no parser disagreements — machine-processing just works.
zlib keeps .rec files ~5–10× smaller on disk and in transit. No effect on context tokens, but real savings on storage and transfer.
Think of .rec as compressed, paginated, checksummed, metadata-rich Markdown. Its real niche is replacing PDF as a transport format — where page fidelity, source provenance, and verifiable integrity matter most.
Know someone paying too much for AI token costs? Share rec-format with them.
Wanna help cover my Anthropic bill?
.rec format saves ~1,500 tokens per page. On a 100-page document, that's 150,000 tokens back in your pocket. Real money on every API call. If it's saving you cost, a small tip keeps this tool alive.