Book-Prep: PDF Book LLM Analysis Preparation Skill

Prepares PDF books in an optimized format for LLM analysis. Provides TOC-based splitting and markdown conversion with AI image descriptions.

Quick Reference

Script Locations

All scripts are located in a single directory within the skill folder:

• Script Directory: .claude/skills/Seongjin_Book-Prep/Scripts/
• Contains: check_toc.py, split_pdf.py, pdf_to_md.py, prompt_book.md, .env

Core Workflow

1. TOC Analysis: Run check_toc.py to generate toc.json
2. Level Selection: Analyze TOC structure to select optimal level
3. PDF Splitting: Run split_pdf.py --level N
4. Markdown Conversion: Convert each split PDF with pdf_to_md.py

Basic Commands

All commands use the Scripts directory. Use absolute path or set SCRIPTS_DIR variable:

SCRIPTS_DIR="/path/to/vault/.claude/skills/Seongjin_Book-Prep/Scripts"

TOC Analysis:

cd "$SCRIPTS_DIR"
source .venv/bin/activate
python3 check_toc.py "/path/to/book.pdf"

PDF Splitting:

python3 "$SCRIPTS_DIR/split_pdf.py" --level 1 "/path/to/book.pdf"

Markdown Conversion:

python3 "$SCRIPTS_DIR/pdf_to_md.py" -p "/path/to/split/chapter.pdf"

Alternative (direct venv python execution):

"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "/path/to/chapter.pdf"

Note: Use -m gemini or -m gpt only when user explicitly requests a specific model.

Output Structure

book.pdf (original)
book/
  toc.json (TOC analysis result)
  1. Chapter One.pdf (split PDF)
  1. Chapter One.md (markdown)
  image-cache_1. Chapter One.json (image cache)
  2. Chapter Two.pdf
  2. Chapter Two.md
  image-cache_2. Chapter Two.json
  images/ (shared image folder)

Environment Setup

Prerequisites

• Python 3.10+ (use python3, not python on macOS)
• Virtual environment in Scripts directory

Initial Setup

SCRIPTS_DIR="/path/to/vault/.claude/skills/Seongjin_Book-Prep/Scripts"
cd "$SCRIPTS_DIR"
python3 -m venv .venv
source .venv/bin/activate
pip install pymupdf pymupdf4llm google-generativeai openai Pillow python-dotenv

API Keys Configuration

Create or edit .env file in Scripts directory:

OPENAI_API_KEY=your-key
OPENAI_MODEL=gpt-4.1-mini
GOOGLE_API_KEY=your-key
GOOGLE_MODEL=gemini-3-flash-preview

Implementation Guide

Step 1: TOC Analysis

Analyzes the PDF's table of contents structure and saves it as JSON.

SCRIPTS_DIR="/path/to/vault/.claude/skills/Seongjin_Book-Prep/Scripts"
"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/check_toc.py" "/path/to/book.pdf"

Generated toc.json structure:

{
  "pdf_name": "Value.pdf",
  "total_pages": 561,
  "summary": {
    "max_level": 4,
    "levels": {
      "level_1": {
        "count": 6,
        "total_pages": 380,
        "avg_pages": 89.7,
        "min_pages": 1,
        "max_pages": 505,
        "total_characters": 770597,
        "avg_characters": 77059.7,
        "min_characters": 0,
        "max_characters": 330635
      },
      "level_2": {
        "count": 5,
        "total_pages": 342,
        "avg_pages": 101.0,
        "min_pages": 30,
        "max_pages": 152,
        "total_characters": 711232,
        "avg_characters": 79025.8,
        "min_characters": 66422,
        "max_characters": 107262
      }
    }
  },
  "toc": [
    {
      "level": 1,
      "title": "Part I",
      "start_page": 20,
      "end_page": 161,
      "page_count": 142,
      "characters": 236255,
      "has_children": true
    }
  ]
}

TOC Fields

Summary Level Statistics

Step 2: Split Level Decision

Split level decision is the core of this skill. Wrong level selection creates chunks that are too large (Context Overflow) or too small (insufficient context).

Decision Rules

Default rule: Use level=1 when it represents the top-level of main body content

Exception rule: Use level=2 when level=1 contains a single entry encompassing the entire body

Judgment Criteria

Use level=2 when ALL of the following conditions are met:

• A specific level_1 entry occupies 70% or more of total pages
• That entry has has_children: true

Use level=1 when:

• level_1 entries are relatively evenly distributed
• Each entry is an independent Part or Section

Real Analysis Examples

Example 1 - Value.pdf (use level=1):

toc.json analysis:

• level_1: Part I (142p), Part II (205p), Part III (140p)
• Each Part occupies 25-37% of total
• Even distribution, so level=1 is appropriate

Example 2 - The Idea of Justice.pdf (use level=2):

toc.json analysis:

• level_1: Preface (6p), Acknowledgements (5p), ... , "The Idea of Justice" (505p)
• "The Idea of Justice" entry occupies 93% (505/543)
• has_children: true
• Using level=1 creates a single 505-page chunk
• Using level=2 properly splits into Introduction, PART ONE~FOUR

Example 3 - Outclassed.pdf (use level=1):

toc.json analysis:

• level_1: Part I, Part II, Part III, Part IV
• Each Part is appropriately sized
• level=1 is appropriate

Step 3: PDF Splitting

Split the PDF with the determined level.

"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/split_pdf.py" --level 2 "/path/to/book.pdf"

Split result:

The Idea of Justice/
  1. Introduction.pdf
  2. PART ONE The Demands of Justice.pdf
  3. PART TWO Forms of Reasoning.pdf
  4. PART THREE The Materials of Justice.pdf
  5. PART FOUR Public Reasoning and Democracy.pdf

Split Options

• --level 1: Top-level TOC (Part level)
• --level 2: Second-level TOC (Chapter level)
• --level 3: Third-level TOC (Section level)

Step 4: Markdown Conversion

Convert each split PDF to markdown.

"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "/path/to/chapter.pdf"

Conversion result (files created in same folder as PDF):

folder/
  Chapter.pdf (input)
  Chapter.md (markdown)
  Chapter.md.backup (backup)
  image-cache_Chapter.json (image cache)
  images/
    Chapter.pdf-0-0.png
    Chapter.pdf-1-0.png

Model Selection

• -m gpt: Use OpenAI GPT (default)
• -m gemini: Use Google Gemini

Advanced Patterns

Batch Processing

Process multiple PDFs sequentially:

for pdf in "/path/to/split/folder"/*.pdf; do
  "$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "$pdf"
done

Parallel Processing

Run multiple conversions in parallel (background jobs):

"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "chapter1.pdf" &
"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "chapter2.pdf" &
"$SCRIPTS_DIR/.venv/bin/python" "$SCRIPTS_DIR/pdf_to_md.py" -p "chapter3.pdf" &
wait

Large Book Processing Strategy

For books over 500 pages:

1. Understand structure with check_toc.py
2. Consider more granular splitting with level_2 or level_3
3. Adjust so each chunk is around 100-200 pages

Prompt Customization

Modify prompt_book.md file to change image description style:

# Available placeholders
{context_before} - Text before image
{context_after} - Text after image
{image_path} - Image path

Caching Utilization

image-cache_[name].json caches image descriptions:

• Located in the output directory alongside markdown file
• Skips already processed images on re-run
• Delete cache and re-run to refresh descriptions

Troubleshooting

Command not found: python

Error: command not found: python

Cause: macOS uses python3 instead of python

Solution: Always use python3 or activate virtual environment first

Module not found: pymupdf

Error: ModuleNotFoundError: No module named 'pymupdf'

Cause: Virtual environment not activated or packages not installed

Solution:

cd "$SCRIPTS_DIR"
source .venv/bin/activate
pip install pymupdf pymupdf4llm google-generativeai openai Pillow python-dotenv

Externally managed environment

Error: externally-managed-environment

Cause: Attempting to install packages outside of virtual environment

Solution: Always use virtual environment pip:

"$SCRIPTS_DIR/.venv/bin/pip" install pymupdf pymupdf4llm google-generativeai openai Pillow python-dotenv

Corrupted Virtual Environment

Error: No such file or directory: .../python3.14

Cause: venv references a Python path that no longer exists (moved or deleted)

Solution: Recreate the virtual environment:

cd "$SCRIPTS_DIR"
rm -rf .venv
python3 -m venv .venv
./.venv/bin/pip install pymupdf pymupdf4llm google-generativeai openai Pillow python-dotenv

PDF Without TOC

Error message: "No TOC found"

Solution:

• Verify TOC exists in PDF viewer
• If no TOC, manual page range specification required
• Add TOC with Adobe Acrobat and retry

Invalid Page Reference

Error message: "Excluded items (invalid page reference)"

Cause: Some TOC entries in PDF reference incorrect page numbers

Solution: Automatically filtered, can be ignored

API Call Failure

Error message: "API call failed"

Check:

• Verify API key in .env file
• Check API quota
• Verify network connection

Missing Image File

Error message: "Image file not found"

Cause: Image extraction failed during PDF conversion

Solution: Adjust DPI setting or change image format

Dependencies

Required packages:

• Python 3.10+
• PyMuPDF 1.26.7+
• pymupdf4llm
• google-generativeai
• openai
• Pillow
• python-dotenv

Installation (within virtual environment):

pip install pymupdf pymupdf4llm google-generativeai openai Pillow python-dotenv

Works Well With

• pdf skill: PDF manipulation and form processing
• pptx skill: Presentation creation
• Seongjin:Knowledge-Architect: Restructure extracted markdown