Private document conversion for Claude

Claude's long context is built for real documents — contracts, board papers, research archives. Which is exactly the material you least want passing through a random converter website on its way there. Convert on your computer instead; only the text you choose reaches Claude.

The privacy chain, kept intact

When you drop a PDF or Word file into a web-based converter, the original — metadata, embedded content, all of it — lands on a third-party server before Claude ever sees a word. Milldown breaks that pattern: conversion and OCR run entirely on-device, you review the Markdown in the app, trim what shouldn't travel, and paste or drag exactly that into Claude. The original never leaves your computer; what Claude receives is a decision, not a side effect.

Why Claude works better on Markdown

• Structure survives where the source has it. Word, PowerPoint and Excel keep their headings, lists and tables, so "summarize section 4" and "compare the two tables" resolve to the right content. A PDF stores positioned text rather than structure, so it arrives as text in reading order.
• Native PDF attachment bills every page twice. Text plus a per-page image — 1,694–1,878 tokens per page, measured with Anthropic’s own token counter. Markdown carries the words alone: 71–84% fewer tokens on our benchmark corpus, re-saved on every chat submission.
• Projects and long chats stay roomy. A lean document leaves the context window for the conversation. For multi-document work, a Context Pack merges a folder into one paste-able file with a table of contents.

The workflow

1. Drop files — or a whole folder — into Milldown.PDF, Word, PowerPoint, Excel, EPUB, web pages; scans are OCR'd on-device with Apple's Vision framework.
2. Clean up and review.One click strips page furniture; the editor lets you cut anything that shouldn't be shared before it goes anywhere.
3. Drag into Claude as a real .md file.Or copy-paste. The token badge shows what the document costs before you send it.

For RAG on Claude

Building retrieval over your documents? Export heading-aware chunks (512–2,048 tokens, Markdown or JSONL) with provenance front matter — source path, timestamp, SHA-256 — so every retrieved chunk stays traceable to its original. The pipeline is in preparing documents for RAG.