I Turned a Book into an AI Agent (book-to-skill)

TL;DR

A trending open-source repo — book-to-skill (MIT, ~14k★) — ingests a PDF you own and turns it into a queryable agent skill: it reads the book, splits it into chapters, and packages chapters/, a cheat-sheet, glossary, patterns, and a skill.md into your skills folder (works in Claude Code, Codex, Pi). The payoff is progressive disclosure: ask about one concept and the agent loads just that chapter, not the whole book — so you spend far fewer tokens and get better answers than dumping the book into a session or one giant markdown file. Pair it with a diagram skill (Excalidraw) and it stops being search and starts making you smarter — a personal, referenceable knowledge base of the books you've actually read.

Key takeaways

• The pattern: PDF → extract + chapter it → skill folder (chapters, cheat-sheet, glossary, patterns, skill.md) → security scan → query on demand.
• Why it's cheap & better: a skill is a markdown file the agent opens only when relevant. Ask about "bias vs. variance" and it opens chapter 3, not the whole book — minimal tokens, sharper answers. Dumping the whole book into a session (or one huge .md) is expensive and worse.
• A skill = an "employee," not a context dump. This is the same progressive-disclosure principle as a thin, routing CLAUDE.md.
• It's more than search. Feed a retrieved concept into a diagram skill (Excalidraw) to draw it — e.g. a learning-curve chart grounded in the actual chapter.
• Knowledge-base angle: ingest several related books and you've built your own topic knowledge base (he compares it to Karpathy's "LLM wikis"). But don't ingest 80 books into 80 skills — "that's a rag job"; make a consolidated knowledge base instead.
• Install hygiene: read any GitHub skill before installing; prefer a local install (project-scoped) over global unless you truly need it everywhere.

---

How it works

flowchart LR
    P["📕 PDF you own"] --> E["Extract text →<br/>group into chapters"]
    E --> F["Skill folder:<br/>chapters · cheat-sheet · glossary<br/>patterns · skill.md"]
    F --> S["🔒 Security scan"]
    S --> Q["Query on demand<br/>(+ draw with a diagram skill)"]

The resulting skill.md is a knowledge base with a topic index — so invoking /machine-learning-yearning and asking a question sends the agent straight to the right chapter markdown.

Why it beats dumping the whole book

flowchart TD
    subgraph BAD["❌ Dump the whole book"]
      B1["Load entire book into session<br/>(or one giant .md)"] --> B2["1000s of tokens *every* prompt<br/>+ worse answers (too big to query)"]
    end
    subgraph GOOD["✅ book-to-skill"]
      G1["Skill opens only the relevant chapter"] --> G2["Minimal tokens<br/>+ precise, grounded answers"]
    end

The rules (important — mostly about copyright)

• Bring a book you actually own. The tool ships with zero book content; you supply the PDF. Don't pirate anything to feed it.
• Keep the output private. The generated skill contains the book's copyrighted ideas — you have the right to use it yourself, not to distribute it. Sharing a skill of a book someone didn't buy is a liability.
• Know when to skip it. Messy PDFs (no headings, odd formatting — like a draft copy) cost a lot of tokens to ingest; not always worth it.
• Use it on books you'll revisit for their frameworks/mental models — not one-off reads.
• Only a handful at a time. Every ingest adds a skill; 80 books = 80 skills polluting your folder → build a single knowledge base instead.

Who it's for

If you want to study a dense book while you work (not in a separate reading session), keep a queryable knowledge base of books you own, or you're tired of rereading a whole book to find the one idea you half-remember — this is the pattern. The deeper idea for my own setup: it's another instance of turn owned knowledge into a progressively-disclosed skill, exactly how a good agent OS should hold context.