Make Claude cite chapter and verse from your actual PDFs

What it does

This is a thin but structured pipeline—essentially a Claude Code skill wrapper around existing extractors like pdftotext, docling, and ebooklib—that ingests technical books and compiles them into a skill directory. It generates a master index of mental models, per-chapter summaries, a glossary, a patterns file, and a cheatsheet. When you query a topic, Claude loads only the relevant chapter files on demand instead of holding the entire book in context.

The interesting bit

Rather than dumping raw text or doing RAG-style retrieval at query time, the tool performs a one-time compile that extracts the author’s actual frameworks, names them, and structures them for reasoning. The README draws a sharp distinction: RAG finds sentence chunks similar to your query; a skill gives you the named frameworks the author spent years building, ready to apply.

Key highlights

• Supports PDF, EPUB, DOCX, TXT, Markdown, HTML, RTF, and MOBI/AZW through a chain of fallback extractors.
• Picks extraction strategy by book type: fast pdftotext for prose, or docling for technical books with tables and code blocks.
• Chapter files stay on disk until requested, so a 400-page book does not burn your token budget on every conversation.
• Output is synthesized, not excerpted: summaries use a practitioner voice (“Use X when Y”) and front-load the most important mental models.
• Particularly useful for niche references, internal documentation, or recent publications that Claude’s training data likely does not include.

Caveats

• The README’s benchmark shows docling takes 164 seconds to process a 103-page technical book on CPU, while pdftotext finishes in 0.1 seconds; preserving tables and code blocks is not free.

Verdict

Developers who want to keep a specific book’s frameworks at their fingertips while coding will find this more useful than a generic RAG setup. If you are looking to search across fifty books at once, the README itself admits that NotebookLM is the better fit.