Book to Claude skill
Convert a PDF book into a Claude skill folder: a SKILL.md index, one Markdown file per chapter, and the prompt that distils them.
Or choose a file. A text-based PDF of any length; nothing is uploaded.
How to use book to claude skill
Reads a PDF in your browser, repairs what the page layout did to the text, splits it at the chapter boundaries, and hands back a zipped skill folder: a SKILL.md index, one Markdown file per chapter, and DISTILL.md, the prompt that turns those chapters into a distilled skill.
- 1Choose a PDF
Drop the file on the panel or use Choose a PDF. Parsing happens in a background thread on your own machine, so a 400-page book takes a few seconds and the page stays responsive while the counter climbs. The file is never sent anywhere, which is also why there is no size limit beyond what your device can hold in memory.
- 2Check the chapters that were found
The list shows every chapter with its page range and a token estimate. Headings are detected from type size first and chapter wording second, so an unusual layout can produce a heading that is really a subheading or a front-matter block you do not want. Untick anything that should stay out — a foreword, an index, a page of endnotes.
- 3Name the skill and download the folder
Skill name becomes the folder and the name in SKILL.md. The description field is the one line Claude reads to decide whether to open the skill at all, so write it as when to use this, not what it is. Download skill folder writes the zip; unpack it into ~/.claude/skills/ and run DISTILL.md from inside it.
When this tool is useful
Why the download is a scaffold rather than a finished skill
A skill earns its place by being faster to consult than the book, and that comes from distilled knowledge: the mental models, the named patterns, the decision rules, the terms the author defines. Producing those means reading and judging, which is a model’s work. What a browser can do is everything around it — pull the text out, repair it, structure it, and write the frontmatter — so this tool does that part completely and leaves a prompt for the part it cannot honestly fake. Run DISTILL.md in Claude Code and the chapter files become a glossary, a pattern catalogue, and a cheatsheet.
The repair work matters more than it sounds. A PDF stores positioned glyphs, not paragraphs, so every line ends where the column ended: sentences arrive chopped into fragments, words are split across hyphens, and the book title and page number are wedged into the text at the top and bottom of every page. Extraction here groups glyphs into lines by their baseline, rejoins the wraps, mends the hyphen splits, and drops the lines that repeat at the page edge — the same fixes as the PDF text cleaner, applied across a whole book instead of a paste.
Two kinds of PDF will disappoint you, and it is better to know which before you build a skill on one. A scanned book has no text layer at all, only pictures of pages; the tool detects this and says so rather than handing back an empty folder, and no browser can OCR it for you. A two-column layout does have text, but its glyphs are drawn in an order that interleaves the columns, so the extraction reads across the gutter. Skim a chapter file before trusting it — DISTILL.md repeats that warning where you will see it again.
Copyright is worth one clear thought before you start. Distilling a book you own into frameworks and takeaways for your own use is ordinary note-taking with a different output format; publishing a folder of its chapters verbatim is not. The chapter files hold real extracted text, so keep the folder local and let the distilling pass replace those passages with your own condensed version before the skill goes anywhere near a shared repository.
Book to Claude skill questions
No. The parser runs inside your browser in a worker thread and the zip is assembled in memory on your machine. There is no server in this tool to receive a file, which is the same reason nothing else on PasteScrub can see your text.
Yes, immediately: it has valid frontmatter, an index, and chapter files Claude will read on demand. It is a skill over raw source text, though, so it is bigger and slower to consult than one whose chapters have been distilled. Running DISTILL.md is what closes that gap.
In ~/.claude/skills/ for a skill you want everywhere, or .claude/skills/ inside a project for one that belongs to that codebase. Claude loads the frontmatter description at startup and opens the body only when a task matches it.
Chapter detection reads type size and heading wording, and some books defeat both: a thesis set entirely in one size, a report with no headings. When no reliable heading appears, the text is split into parts of a readable length at page boundaries instead, which is why a chapter may be called Part 3 with a page range.
Not yet — PDF only, and only a PDF that carries a text layer. A scanned book is detected and refused rather than silently producing an empty skill. For other formats, book-to-skill by virgiliojr94 runs the same idea as a Claude Code skill with Python extraction behind it.
Their labels are text on the page like anything else, so they extract — but a chart’s axis ticks and data labels sit at many different heights and interleave with the prose when the page is read top to bottom. Rather than leave an axis running through a paragraph, each chapter file collects that text under a Figure and table text heading, grouped by page, with the caption left in the prose where the author put it. The numbers survive; the axis they belonged to does not. For a book that is mostly figures, Docling by IBM Research does the job properly: it recognises table structure and converts bar, pie, and line charts into tables with descriptions. It is a Python tool with model weights, so it cannot run in a browser tab.
The chapter list shows an estimate per chapter and a total before you download, counted at four characters per token. Only the index loads by default; a chapter costs its own estimate when Claude opens it, which is the point of splitting the book up in the first place.