Turning Technical Books into Claude Code Skills: Unlocking Internal Documentation as Reusable Skills

Tool Overview

Give a PDF or EPUB to ~/.claude/skills/<slug>/. The tool generates structured files: SKILL.md (core framework + chapter index) and per‑chapter markdown files, plus a glossary, patterns list, and cheatsheet.

On‑Demand Loading

A 400‑page technical book is roughly 200 K tokens. Loading the whole text each conversation is costly. The skill loads only the ~4 K‑token SKILL.md by default; a chapter file (~1 K tokens) is read only when a specific topic is queried, dramatically reducing token consumption.

Difference from Direct PDF Ingestion

Direct PDF ingestion performs keyword retrieval and returns page ranges (e.g., “pages 117‑135 mention replication”). The skill extracts the author’s framework, naming conventions, and mental models, so a query about data replication returns a concise summary of the three replication models and their applicable scenarios.

Not a Retrieval‑Augmented Generation (RAG) System

RAG searches across many documents at query time. The skill compiles a single book once, extracting its structure, making it ideal for deep, frequent reference to one source.

Technical Implementation

The project consists of two files: SKILL.md – skill definition, ≈530 lines. scripts/extract.py – text extraction script, ≈830 lines.

Supported source formats: PDF, EPUB, DOCX, TXT, Markdown, reStructuredText, AsciiDoc, HTML, RTF, MOBI/AZW/AZW3.

PDF extraction has two modes:

text mode : pdftotext → PyPDF2 → pdfminer, finishes in seconds, suitable for plain‑text books.

technical mode : uses Docling to preserve tables and code blocks; for a 103‑page book it takes about 164 s and retains 48 tables and 36 code blocks.

Benchmark on a 103‑page technical book:

pdftotext – 0.1 s, 27 K tokens, no tables, no code.

Docling – 164 s, 27 K (+1.2 %) tokens, 48 tables, 36 code blocks.

Compilation Workflow

Document file
  │
  ▼
User selects: technical book or plain‑text book
  │
  ├── technical → Docling (preserve tables & code)
  └── plain‑text → pdftotext chain (fast)
  │
  ▼
extract.py outputs full text + metadata
  │
  ▼
Claude analyzes structure (title, author, chapters, TOC)
  │
  ▼
Generate per‑chapter summaries (800–1,200 tokens each)
  │
  ▼
Add glossary, patterns, cheatsheet
  │
  ▼
Create SKILL.md (core framework + index)
  │
  ▼
Write to ~/.claude/skills/<slug>/

For books larger than 50 K tokens (≈130 pages), the tool slices the book by chapter using grep + sed, loading only the needed chapter and saving an order of magnitude in token cost.

Installation

mkdir -p ~/.claude/skills/book-to-skill/scripts
curl -o ~/.claude/skills/book-to-skill/SKILL.md https://raw.githubusercontent.com/virgiliojr94/book-to-skill/master/SKILL.md
curl -o ~/.claude/skills/book-to-skill/scripts/extract.py https://raw.githubusercontent.com/virgiliojr94/book-to-skill/master/scripts/extract.py

Running the Compiler

# PDF – auto‑derive skill name
/book-to-skill ~/Downloads/designing-data-intensive-applications.pdf

# EPUB – specify custom name
/book-to-skill ~/books/clean-code.epub clean-code

# Full path and custom name
/book-to-skill /tmp/ddd-evans.pdf domain-driven-design

Using the Skill

/designing-data-intensive-apps          # load core framework
/designing-data-intensive-apps replication   # query a specific topic
/designing-data-intensive-apps ch05          # open chapter 5

Dependencies

Plain‑text books: pdftotext (poppler‑utils).

Technical books: pip3 install docling.

EPUB handling (optional): pip3 install ebooklib beautifulsoup4.

Design Principles

Density over completeness : a 1 000‑token summary is more useful than a 10 000‑token verbatim excerpt.

Practitioner perspective : output statements such as “use Y when X” instead of merely citing chapter locations.

Pre‑load SKILL.md : place the most important content at the beginning because Claude Code truncates from the end.

Never copy raw text : chapter files are distilled, preserving copyright and information density.

Token‑cost estimate : the compiler reports input/output token counts and estimated monetary cost before proceeding.

Suitable Scenarios

Repeatedly referenced technical books (e.g., DDIA, Clean Code, DDD).

Internal company documentation not present in Claude’s training data.

Newly published books that Claude has not yet ingested.

Non‑English technical books with limited training coverage.

Unsuitable Scenarios

Search across dozens of books – better suited for RAG.

Books Claude already knows well (e.g., Python Cookbook) – limited benefit.

Scanned PDFs without a text layer – require OCR first.

Limitations

The tool’s ceiling depends on Claude’s extraction quality; deep logical chains or heavy context dependencies may not be fully captured in a 1 000‑token chapter summary, though most framework‑oriented books work well.

Conclusion

book-to-skill extracts a book’s core structure, splits it by chapter, and installs it as a Claude Code skill. Compared with raw PDF ingestion, it offers on‑demand loading, predictable token usage, and more reliable answers. The project consists of two files (≈1 300 lines total) and has over 2 300 GitHub stars.

GitHub repository: https://github.com/virgiliojr94/book-to-skill