pdf-to-markdown

PDF to Markdown Converter

Extract complete PDF content as structured Markdown, preserving:

• Headers (detected by font size, converted to # tags)
• Bold, italic, monospace formatting
• Tables (converted to Markdown tables)
• Lists (ordered and unordered)
• Multi-column layouts (correct reading order)
• Code blocks
• Images (extracted and copied next to output with relative paths)

When to Use This Skill

USE THIS when:

• User wants the "whole PDF" or "entire document" in context
• Analyzing, summarizing, or discussing PDF content
• User says "load", "read", "bring in", "extract" a PDF
• Grepping/searching would miss context or structure
• PDF has tables, formatting, or structure to preserve

Environment Setup

This skill uses a dedicated virtual environment at ~/.claude/skills/pdf-to-markdown/.venv/ to avoid polluting the user's working directory.

First-Time Setup (if .venv doesn't exist)

# For fast mode only (PyMuPDF):
cd ~/.claude/skills/pdf-to-markdown && uv venv .venv && uv pip install --python .venv/bin/python pymupdf pymupdf4llm

# For --docling mode (high-accuracy tables):
cd ~/.claude/skills/pdf-to-markdown && uv venv .venv && uv pip install --python .venv/bin/python pymupdf docling docling-core

# Or install everything:
cd ~/.claude/skills/pdf-to-markdown && uv venv .venv && uv pip install --python .venv/bin/python pymupdf pymupdf4llm docling docling-core

Verify Installation

# Verify fast mode:
~/.claude/skills/pdf-to-markdown/.venv/bin/python -c "import pymupdf; import pymupdf4llm; print('OK')"

# Verify docling mode:
~/.claude/skills/pdf-to-markdown/.venv/bin/python -c "import pymupdf; import docling; import docling_core; print('OK')"

Quick Start

# Convert PDF to markdown (always extracts images)
~/.claude/skills/pdf-to-markdown/.venv/bin/python ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py document.pdf

# Output: document.md + images/ folder (next to the .md file)

Standard Workflow

When user provides a PDF and wants full content in context:

Step 1: Ensure the skill venv exists

test -d ~/.claude/skills/pdf-to-markdown/.venv || (cd ~/.claude/skills/pdf-to-markdown && uv venv .venv && uv pip install --python .venv/bin/python pymupdf pymupdf4llm)

Step 2: Convert PDF to Markdown

~/.claude/skills/pdf-to-markdown/.venv/bin/python ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py "/path/to/document.pdf"

Step 3: Read the output

# Output is written to document.md in the same directory as the PDF
cat /path/to/document.md

Caching

PDFs are aggressively cached to avoid re-processing. First extraction is slow, every subsequent request is instant.

How It Works

• Cache location: ~/.cache/pdf-to-markdown/<cache_key>/
• Cache key: Based on file content hash + extraction mode
• Invalidation: Cache is invalidated when:
  • Source PDF is modified (size or mtime changes)
  • Extractor version changes (automatic re-extraction)
  • Explicitly cleared with --clear-cache or --clear-all-cache

Cache Commands

# Clear cache for a specific PDF
~/.claude/skills/pdf-to-markdown/.venv/bin/python ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py document.pdf --clear-cache

# Clear entire cache
~/.claude/skills/pdf-to-markdown/.venv/bin/python ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py --clear-all-cache

# Show cache statistics
~/.claude/skills/pdf-to-markdown/.venv/bin/python ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py --cache-stats

Cache Contents

~/.cache/pdf-to-markdown/<cache_key>/
├── metadata.json    # source path, mtime, size, total_pages
├── full_output.md   # cached full markdown
└── images/          # extracted images

Image Handling

Images are always extracted. They are:

1. Cached in ~/.cache/pdf-to-markdown/<cache_key>/images/
2. Copied to images/ folder next to the output .md file
3. Referenced in the markdown with relative paths (images/filename.png)
4. Summarized in a table at the end of the document

Auto-View Behavior for Images

IMPORTANT: When the extracted markdown contains image references like:

**[Image: figure_1.png (1200x800, 125.3KB)]**

And the user asks about something that might be visual (charts, graphs, diagrams, figures, screenshots, layouts, designs, plots, illustrations), automatically use the Read tool to view the relevant image file(s) before answering. Don't ask the user - just look at it.

Examples of when to auto-view images:

• User: "What does the chart on page 3 show?" → Read the image file
• User: "Summarize the figures in this paper" → Read all image files
• User: "What's in the diagram?" → Read the image file
• User: "Describe the architecture shown" → Read the image file
• User: "What are the results?" (and there's a results figure) → Read it

Output Format

The markdown output includes:

Header (metadata)

---
source: document.pdf
total_pages: 42
extracted_at: 2025-01-15T10:30:00
from_cache: true
images_dir: images
---

Content with image references

# Main Title

## Section Header

Regular paragraph text with **bold**, *italic*, and `code` formatting.

![Figure 1](images/figure_1.png)

**[Image: figure_1.png (800x600, 45.2KB)]**

| Column A | Column B |
|----------|----------|
| Data 1   | Data 2   |

Image summary table (at end)

---

## Extracted Images

| # | File | Dimensions | Size |
|---|------|------------|------|
| 1 | figure_1.png | 800x600 | 45.2KB |
| 2 | chart_2.png | 1200x800 | 89.1KB |

Script Reference

Location: ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py

Usage: pdf_to_md.py <input.pdf> [output.md] [options]

Options:
  --docling         Use Docling AI for high-accuracy tables (~1 sec/page)
  --no-progress     Disable progress indicator

Cache Options:
  --clear-cache        Clear cache for this PDF and re-extract
  --clear-all-cache    Clear entire cache directory and exit
  --cache-stats        Show cache statistics and exit

High-Accuracy Mode (Docling)

For PDFs with complex tables that need high accuracy, use the --docling flag:

~/.claude/skills/pdf-to-markdown/.venv/bin/python \
    ~/.claude/skills/pdf-to-markdown/scripts/pdf_to_md.py \
    document.pdf --docling

When to use --docling:

• PDF has complex tables (borderless, merged cells, multi-column)
• Table accuracy is critical (medical data, financial reports)
• You're seeing garbled table output in default mode

Trade-offs:

• ~1 second per page (vs instant for fast mode)
• First run downloads AI models (~500MB one-time)
• Higher-resolution images (4x default)

Note: --accurate is an alias for --docling.

Troubleshooting

"No module named pymupdf4llm" or venv doesn't exist

Recreate the skill's virtual environment:

# For fast mode:
cd ~/.claude/skills/pdf-to-markdown && rm -rf .venv && uv venv .venv && uv pip install --python .venv/bin/python pymupdf pymupdf4llm

# For docling mode:
cd ~/.claude/skills/pdf-to-markdown && rm -rf .venv && uv venv .venv && uv pip install --python .venv/bin/python pymupdf docling docling-core

Poor extraction quality

• Try --docling for complex tables
• For scanned PDFs, ensure Tesseract OCR is installed: brew install tesseract

Tables not formatting correctly

For complex tables, use --docling mode which uses IBM's TableFormer AI model.

Comparison with Other Approaches

Approach	Use Case	Limitations
This skill (pymupdf4llm)	Full document context with images	Large PDFs may exceed context
--docling mode	Complex tables, medical/financial PDFs	Slower (~1 sec/page), larger models
Grepping PDF	Find specific text	Loses structure, no images
Page-by-page extraction	Targeted pages	Manual, loses cross-page context
Read tool on PDF	Quick preview	Limited formatting preservation