Forgetting a book three months after reading it isn't your fault — this open-source tool compiles entire books into an AI's brain

Have you ever spent $30 on a technical book, devoured the first chapter and highlighted half the pages, only to be asked three months later what it was about and manage a vague "I think it was about... databases? It was pretty good"? That's not a bad memory on your part. A 400-page technical book contains roughly 200,000 characters; reading it once and expecting to remember every chapter months later simply isn't realistic. What if there were a way to make the knowledge in a book available on demand — no memorization, no forgetting?

💡 What You Will Learn

Have you ever spent $30 on a technical book, devoured the first chapter and highlighted half the pages, only to be asked three months later what it was about and manage a vague "I think it was about..

📜 Table of Contents

• Have you ever had this experience?
• The traditional approach doesn't work
• A different approach: compile once, reuse forever
• What does a book get broken into?
• How to use it? Up and running in a minute
• What formats are supported?
• What's the background of this project?
• Who needs it most?
• In one sentence

Have you ever had this experience?

You spend a hundred-odd yuan on a technical book, devour the first chapter with feverish excitement, and cover the pages in highlights. Three months later someone asks what the book was about, and after racking your brain you blurt out, "I think it covered... databases or something. It was pretty impressive."

This isn't a bad memory on your part.

A 400-page technical book, 200,000 characters — finishing it front to back is already an achievement. Expecting to still recall what every chapter said three months later was never realistic.

So is there a way to make the knowledge in a book always available on demand, without memorizing, without forgetting?

There actually is.

The traditional approach doesn't work

First, let's talk about how most people "read with AI."

Approach one: throw the entire PDF at the AI. What happens? The AI's context window instantly fills with hundreds of pages. You ask one question, it starts answering from page 50, reaches page 300 and has already forgotten what it said earlier. Every question burns tokens — one or two questions is fine, but ask too many and costs skyrocket.

Approach two: build a RAG retrieval system. Sounds professional, but it's essentially keyword matching. The AI pulls out every place your book mentions "replication," stitches them together, and calls it an answer. Is the logic sound? Depends on luck. Are the ideas connected? Nope.

Both approaches share the same fatal flaw — every question reprocesses the whole text. It's cramming at the last minute, and even the gods are tired of it.

A different approach: compile once, reuse forever

There's an open-source project on GitHub called book-to-skill that flips the whole idea around.

It doesn't do on-the-fly retrieval, and it doesn't stitch text together. Instead, it thoroughly digests, breaks down, and structures the entire book in advance, compiling it into a Skill file the AI can call directly. Compile once, then ask anything any time — load on demand, never waste a single token.

Think of it this way: you're not making the AI flip through the book; you're loading the book into the AI's brain.

What does a book get broken into?

Run /book-to-skill ./your-book.pdf and it generates a complete Skill package:

A 400-page technical book is roughly 200k tokens in full. But for day-to-day use, the AI only loads the core ~4000-token framework plus the chapter you're asking about.

Official measured data: answering one question saves 24 to 51 times the tokens compared to dumping the whole book in.

How to use it? Up and running in a minute

Installation follows the AI-native route — type a single line into Claude Code or Copilot CLI:

Install book-to-skill:https://raw.githubusercontent.com/virgiliojr94/book-to-skill/master/SKILL.md

Then point it at your ebook and run:

/book-to-skill ./DDIA.pdf

Once compiled, just ask whatever you want to know:

/DDIA replication consistency

The AI locates the relevant chapter itself and answers from the real content. No fabricating, no "see page X," no listing page numbers.

What formats are supported?

It handles 9 document formats:

• Ebooks: PDF, EPUB, MOBI / AZW (Kindle format)
• Office documents: DOCX, TXT
• Markup languages: Markdown, HTML, RTF

Note: scanned PDFs and encrypted documents aren't supported — you'll need to OCR or decrypt them first.

For technical books it uses the Docling engine, fully preserving tables, code blocks, and layout. For purely narrative books it uses pdftotext for blazing-fast extraction.

What's the background of this project?

MIT open-source license, by author virgiliojr94. Currently 83 commits and 14 contributors; it has reached #10 on the Trendshift Python daily chart and #25 on the overall site ranking.

The code is lightweight: the core is just two files — SKILL.md (~530 lines) and extract.py (~830 lines) — and it runs on a laptop with 8GB of RAM.

Who needs it most?

• Frequent book-flippers — classics like DDIA, Clean Code, SICP to consult anytime while coding, 10x faster than flipping through the physical book
• Internal corporate documentation users — architecture manuals, ops runbooks, onboarding docs that can never be in AI training data
• New-book early adopters — content too recent for AI to have trained on, injected straight into your workflow
• Notes and organization nerds — consolidating Markdown notes and scattered documents into one unified Skill, set once and done

In one sentence

Feed the PDF directly and the AI is a page-turner telling you "see page X"; RAG makes the AI a copy-paster stitching things at random; book-to-skill makes the AI a learner that has internalized the whole book's structure and answers whatever you ask.

One is cramming at the last minute; the other is growing a brain permanently.

Which do you pick?

Source: GitHub / Blog Garden (博客园)