book-to-skill is a converter that turns technical books you already own into agent skills. Its GitHub repository states the goal plainly: “Turn any technical book, document folder, or collection of sources into a unified agent skill — ready to study, reference, and use while you work.”
The tool is written in Python, released under the MIT licence, and produces skills that work with Claude Code, GitHub Copilot CLI, and Amp through the open Agent Skills format.
The problem it addresses
The premise is familiar to anyone with a shelf of technical books: you read one carefully, and six months later you remember that it covered replication strategies well, but not what it actually said. The usual workarounds are unsatisfying. Dumping a PDF into a prompt is expensive and imprecise. Notes get written once and never opened again.
book-to-skill’s answer is to restructure the book into something an agent can navigate. Instead of one large blob, you get a small always-loaded core plus per-chapter files that load only when relevant.
What it produces
Running the converter creates a skill directory with a documented shape and approximate token budget:
SKILL.mdat roughly 4,000 tokens, holding the core frameworks and a chapter index.chapters/files at roughly 1,000 tokens each, loaded on demand.glossary.md(~1,500 tokens),patterns.md(~2,000 tokens), andcheatsheet.md(~1,000 tokens) as reference material.
The extraction step pulls out frameworks, decision rules, patterns, and a glossary rather than simply splitting text at chapter boundaries. Input can be a single file, a folder, a glob, or a list, and the documented formats cover PDF, EPUB, DOCX, Markdown, HTML, RTF, MOBI, and plain text, using format-specific parsers such as pdftotext, pypdf, pdfminer.six, Docling, and ebooklib.
Using it
Installation is a clone into your agent’s skills directory:
git clone https://github.com/virgiliojr94/book-to-skill.git ~/.claude/skills/book-to-skillThe path changes per agent — ~/.copilot/skills/ for Copilot CLI, ~/.agents/skills/ for Amp or cross-agent setups. Conversion is then a single command:
/book-to-skill <path|folder|glob> [skill-name]Afterwards the book itself becomes a command. As the README puts it: “Once installed, you just type /your-book-slug replication and the agent reads the right chapter and answers from the actual content.”
Reading the token numbers properly
The headline claim is “24×–51× fewer tokens than dumping the book into context.” That number is real but worth understanding before you repeat it, because the project’s own performance documentation is more careful than the headline.
The measurements use tiktoken with cl100k_base across four books ranging from 244 to 501 pages — Think Python 2 at 119K tokens, Working Backwards at 175K, Moby-Dick at 301K. A targeted question is modelled as the resident core (~4K tokens) plus one chapter (~1K), around 5,000 tokens in total, which is where the ratio against a full-book context dump comes from.
Two qualifications matter. First, the documentation separates that figure from a second, smaller one: a discovery-loop advantage of “2.4–15.6×”, reflecting the up-front cost of building the skill. Second, and more usefully, it explains why the big number is the meaningful one: “The context-dump advantage (24–51×) is the strongest claim: that cost recurs on every conversation turn,” while the build cost is one-time.
So the honest framing is that you pay a one-off conversion cost to avoid re-sending an entire book on every turn. That is a sound trade, and the project deserves credit for documenting the weaker ratio alongside the impressive one instead of quietly dropping it.
Copyright, handled carefully
This is the part that could easily have been glossed over, and is not. The repository ships no book content at all — “not a single page. It’s a converter you point at files you already own.” Processing is local: “Extraction and analysis run on your machine. Your files are never uploaded by this tool.”
The licensing boundary is drawn just as explicitly. MIT covers the converter code, not the documents you feed it or the skills it produces. The documentation treats generated skills as personal study notes and advises against redistributing them where they contain copyrighted material.
That distinction is worth internalising rather than skimming. A generated skill is a dense structured summary of a book — exactly the kind of artefact that feels shareable and should not be. Keep them local, and the tool sits comfortably within normal personal use of a book you bought.
Practical considerations
- Output quality tracks the source. A well-structured technical book with clear chapters converts better than a scanned PDF with no usable text layer. Check what the extractor produced before trusting it.
- The parser stack has moving parts. Several optional extractors are supported for PDFs; expect to install dependencies for the formats you actually use, and check the CLI’s dependency check.
- Token budgets are estimates. The per-file numbers describe typical output, not guarantees for a dense 900-page reference.
Verdict
book-to-skill is a focused tool with an unusually honest presentation: local processing, no bundled content, a clear licence boundary, and performance documentation that qualifies its own headline. If you work alongside an agent CLI and own technical books you would consult more often if consulting them were cheaper, the setup cost is one clone and one command per book.
