Say Goodbye to “Read and Forget”: How book-to-skill Turns Knowledge into AI’s “Permanent Skills”#

Distill an entire book into a skill your AI can call on demand—rather than text waiting to be retrieved

Core Insight: It’s Not Summarization, It’s “Compilation”#

The one-line positioning of book-to-skill is this: take any technical book, documentation folder, or collection of materials, and turn it into a unified Agent Skill that you can consult while working inside Claude Code, GitHub Copilot CLI, or similar tools.

It works in just three steps, simple enough to state in one sentence:

Point → It distills → Agent loads on demand

The key is that “distillation” is not “summarization”. Summarization is compression; distillation is extracting frameworks, decision rules, anti-patterns, and mental models—encoding the knowledge system that an author spent years building into a structure that AI can directly reason with, rather than a pile of text waiting to be keyword-matched.

The difference is obvious in daily use: when you ask “how to do data replication in distributed systems,” you get “this book covers 3 replication models, each suited for specific scenarios,” not “pages 117 to 135 mention replication.”

The Token-Saving Secret: Eliminating the “Discovery Loop Tax”#

Why is throwing a PDF directly at AI both expensive and ineffective? The author coins a vivid term for it: the “Discovery Loop Tax”.

An agent reading a PDF doesn’t just “read”—it also has to “navigate”: you ask a question, it grabs the table of contents, spots an unfamiliar term, flips a few pages, backtracks… each jump goes into the conversation history and gets reprocessed in every subsequent turn. To stay within budget, sub-agents are forced to brutally compress, eventually handing you a low-quality summary you can’t verify.

book-to-skill’s approach is: the navigation cost is paid only once, at compile time. At runtime, it loads only a compact resident core + one pre-compiled chapter (about 5,000 tokens). No discovery loop, no compression distortion, and the full original text remains on disk for verification.

Official Numbers from Three Real Books#

Dropping a book into context is paying tax “every round, every session”; book-to-skill is paying a one-time compilation fee, then fetching only the slices you need. In economic terms, this is called “amortization”.

What It Generates: A Layered, On-Demand Knowledge System#

Run /book-to-skill your-book.pdf once, and it produces a complete set of structured files inside your skills directory:

The design essence is on-demand loading: normally only the 4,000-token core skeleton is resident; specific chapters are fetched only when a topic comes up. A 1,000-token distillation beats a 10,000-token raw excerpt—density over completeness.

Don’t Be Fooled by the “Book” in Its Name: It Can Digest Far More#

It’s called book-to-skill, but the input is actually any structured text:

The litmus test is simple:

If you find yourself repeatedly opening a document and wishing you had it memorized, it’s good raw material.

The transfer value of this idea may be greater than the tool itself: instead of stuffing long materials into the dialog box whole, first distill them into structured, on-demand knowledge.

Hands-On: 5 Steps to Get Started#

Step 1: Install Dependencies#

# Run a health check first
python3 scripts/extract.py --check

# For plain-text books
sudo apt install poppler-utils

# For technical books (with code, tables)
pip3 install docling

For technical books, Docling’s layout-aware extraction tells the whole story—a 103-page technical book comparison:

In technical books, tables and code are critical—those extra two minutes are well worth it.

Step 2: Install the Skill#

Inside Claude Code, simply say:

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

Step 3: Point It to Your Book#

/book-to-skill ~/Downloads/designing-data-intensive-applications.pdf
/book-to-skill ~/books/clean-code.epub clean-code

Step 4: Answer “Technical or Text-Only Book?”#

The tool will ask this question, then automatically select the right parsing engine. Don’t skip it—choosing wrong will lose tables and code.

Step 5: Use It Like a Skill#

/designing-data-intensive-apps                # Load the core framework
/designing-data-intensive-apps replication    # Look up a specific topic
/designing-data-intensive-apps ch05           # Jump directly to Chapter 5

Costs and Limitations#

Costs#

Compiling a book costs roughly $1 USD (estimated with Claude Sonnet 4.5), after which each use consumes only a few thousand tokens.

Limitations#

1. Requires clear chapter markers (“Chapter N” works best)
2. Scanned PDFs without a text layer—need OCR first
3. If a book’s logical chain is exceptionally deep or context-dependent, a 1,000-token chapter summary may not fully capture it. But for most “framework-style” technical books—those about patterns, principles, and decision trees—it works remarkably well

Conclusion#

What book-to-skill does is not complicated: extract the core structure of a technical book, split it by chapters, and package it into AI’s skill system. But this approach is far more refined than “dumping a PDF into context”—compile once, load on demand, keep token consumption under control, and get more stable answer quality.

Pull that technical book off your shelf—the one you read once and left to gather dust—run the environment check, and turn it into an “external brain” you can call upon anytime. Compiling a book costs about $1 USD, far cheaper than re-feeding the entire PDF in every session.