Stop re-reading that tech book — book-to-skill distills any PDF into an on-demand Claude Code skill (and saves 24–51× tokens)

What it does

book-to-skill does one thing really well: it turns any technical book, document folder, or collection of sources you own into a unified agent skill that Claude Code (and also GitHub Copilot CLI and Amp) can load on demand. The key idea is that it does not dump the whole book into your context window and burn tokens. Instead, it first distills the book into a structured knowledge skeleton — core mental models, per-chapter notes, a glossary, design patterns, and a cheatsheet — and only when you are actually working do you type a single command like /your-book-slug replication and the agent goes and reads the exact chapter it needs, answering from the real source content without hallucinating.

Why it is blowing up

It rides two of the hottest waves right now. First, the Agent Skills open standard pushed by Anthropic, where a single SKILL.md file is understood by many different agents. Second, the universal pain of "the AI forgets what it read." You spend money on a great book, read it once, and three months later you cannot even remember chapter seven existed. Searching the PDF gives you a list of page numbers, not answers. Asking your agent about the book makes it hallucinate or claim it has no content. book-to-skill breaks that loop with one command. The repo was created on 2026-08-13 and crossed 1,000 stars within a single day — a textbook "someone should have built this years ago" tool.

Technical highlights

What it generates is not a flat summary but a whole set of on-demand files. SKILL.md (~4,000 tokens) holds the core mental models plus a chapter index and acts as the agent's master outline. chapters/ch01-*.md keeps one file per chapter, loaded only when needed, so unasked chapters never eat into the token budget. glossary.md lists key terms with references to the chapters where they appear. patterns.md collects techniques, algorithms, and design patterns. cheatsheet.md gives decision tables and quick-reference rules. Under the hood it has two halves: a deterministic Python extractor (PDF, EPUB, DOCX, MD, HTML, RTF, MOBI → clean text plus metadata) and a spec-driven generator (the agent follows SKILL.md to turn that text into a structured skill). The crucial piece is on-demand loading — the chapter files normally do not count against the skill budget, and the agent only reads the chapter you actually ask about. Measured on real books, answering a single question costs 24× to 51× fewer tokens than pasting the entire book into context.

Beyond books

The name says "book," but the input is any structured prose. Internal documentation such as ADRs, runbooks, and onboarding guides is a natural fit. Brand and design systems — voice guidelines, component principles — turn into a reference the agent can cite. Research clusters of papers plus notes can be folded in as new material lands. Specs and standards like RFCs, API contracts, and compliance docs become queryable. The rule of thumb: if you reopen a document often enough to wish you had memorized it, it is a candidate.

Who it is for

• Students and self-learners: turn a textbook into a conversational tutor you can grill before an exam — ask "how is that formula in chapter seven derived" and it opens the real content.
• Developers: turn framework docs, API contracts, and RFCs into a living dictionary on your workbench.
• Research teams: cluster papers plus notes into one skill, and fold in new material as it arrives.
• Companies: ADRs, runbooks, onboarding guides, and brand guidelines shared as one knowledge skill that never goes stale.

Quick try

pip install -e .
book-to-skill install        # copies the skill into ~/.claude/skills and friends
book-to-skill ./my-book.pdf  # distill in one shot

It also supports book-to-skill "docs/*.epub" my-slug for batches and --check to self-test dependencies. The skill installs into ~/.claude/skills, ~/.copilot/skills, and ~/.agents/skills (and related paths) so multiple hosts share one format. On the first two CLI runs a small GUI is unpacked and launched automatically; force it later with book-to-skill ui. Text-heavy PDFs use poppler or pypdf; technical books with code and tables are best handled with docling.

Versus the alternatives

It is different from note-taking apps and from RAG question answering. RAG chunks documents and retrieves by vector similarity; book-to-skill distills the book into a structured knowledge skeleton where the agent receives "what this chapter is about, which decision rules apply, which anti-patterns to avoid" rather than a pile of text fragments. It runs fully locally, never uploads your files, is MIT-licensed, and ships no copyrighted content — you point it at books you already own, and the output is structured notes, not a reproduction, so you should not redistribute generated skills of copyrighted works.