doc2md: Solving PDF Reading in Claude Code (15K lines, multi-extractor, QC pipeline)

[neoncapy]

The Problem

Anthropic's copyright filter blocks most PDF content in Claude Code. If you work with academic papers, regulatory documents, or corporate presentations, you have likely hit this wall. Your agent reads a PDF and gets partial or empty results.

doc2md

doc2md is a 15,000-line Python pipeline that converts PDF, DOCX, and PPTX files into high-fidelity Markdown with full image extraction and multi-stage quality control.

How It Works

2-tier architecture:

1. Python tier (zero LLM tokens) — extracts text, tables, and images using multiple extractors
2. Claude vision tier — 8 expert personas analyze extracted images with document-aware context

Multi-extractor support:

• pymupdf4llm (default)
• pdfplumber (cross-validation)
• MinerU (complex layouts with tables/figures)
• Automatic fallback when one extractor struggles

Per-image classification:
8 heuristics classify every image as substantive or decorative — file size, pixel variance, aspect ratio, color count, vector content detection, journal branding patterns, near-black detection.

QC pipeline:
Structural QC catches table collapse, missing headings, dropped content. Loops until genuinely zero issues remain.

Claude Code Integration

• PreToolUse hook intercepts PDF/DOCX/PPTX reads and redirects to converted Markdown
• SHA-256 conversion registry tracks every file
• SKILL.md provides the full pipeline as a slash command

What It Handles

• Scientific papers with complex tables and figures
• Pharmaceutical regulatory documents
• Corporate PPTX presentations with charts and SmartArt
• DOCX files with merged cells, images, and nested tables

MIT licensed. Feedback welcome — what document types cause you the most problems?

https://github.com/orangefineblue/doc2md

[hesreallyhim]

Please don't open any more self-promotion stuff.