youtube-to-book

Turn a YouTube video or playlist into a real book — a typeset LaTeX PDF with chapters, syntax-highlighted code, callout boxes, and hand-authored TikZ diagrams.

It is a Claude Code skill. You install it once, paste a URL, and Claude does the rest: pulls the transcript, grabs the frames worth keeping, outlines the book, writes each chapter in parallel, reviews the prose, looks at every rendered diagram to check it's actually correct, and compiles the PDF.

you:    turn this into a textbook: https://www.youtube.com/watch?v=...
claude: ✓ ./textbooks/database-internals/main.pdf — 9 chapters, ~113 pages

Install

npx youtube-to-book install

That copies the skill into ~/.claude/skills/youtube-to-book/ and checks your dependencies. Restart Claude Code and it's live.

Project-local instead of global:

npx youtube-to-book install --project   # → ./.claude/skills/youtube-to-book/

Other commands:

| Command | What it does | |---|---| | npx youtube-to-book install | Install the skill (--force to overwrite, --dir <path> for a custom location) | | npx youtube-to-book doctor | Check that the external tools are present | | npx youtube-to-book where | Print the install path | | npx youtube-to-book uninstall | Remove it |

Dependencies

The skill shells out to these. doctor tells you which are missing.

brew install yt-dlp ffmpeg tectonic poppler

| Tool | Used for | |---|---| | yt-dlp | transcripts, metadata, media URLs | | ffmpeg | key-frame extraction | | tectonic | LaTeX compilation (auto-fetches packages, no TeX Live install) | | python3 | the figure-render script | | pdftoppm (poppler), or macOS sips, or ImageMagick | rasterising figures for review |

Usage

Just talk to Claude Code:

• "make a book from this video: <url>"
• "turn this playlist into a textbook" — each video becomes a chapter
• "convert this to a book, but go deeper on the storage-engine parts"

Output lands in ./textbooks/<book-slug>/ relative to your working directory. Set TEXTBOOKS_DIR to change the default, or just name a directory in your request.

textbooks/<book-slug>/
├── main.tex          preamble, title page, TOC
├── chapters/         ch01.tex, ch02.tex, ...
├── frames/           key frames that survived review
├── figrender/        every TikZ figure as a PNG, for the visual review pass
├── transcript.txt
├── outline.md
└── main.pdf          ← the book

How it works

1. Extract — transcript via yt-dlp, cleaned of VTT timestamps and dedup noise. Frames are pulled targeted, not swept: chapter markers or transcript cues pick 10–20 moments, and ffmpeg -ss before -i seeks by HTTP range so each frame costs seconds instead of streaming the whole video.
2. Outline — chapters, key concepts, which figures belong where.
3. Write — one subagent per chapter, in parallel, each given only its own transcript slice and an explicit list of TikZ figures to author.
4. Review — two independent passes, both load-bearing:
   • a prose pass for cross-chapter repetition, factual drift, and tone breaks;
   • a visual pass that renders every tikzpicture to PNG and looks at it.
5. Compile — self-correcting tectonic loop until the PDF builds.

Why the visual review exists

A geometrically wrong diagram compiles with zero warnings and reads fine in source.

Real case: the four SQL JOIN types were drawn as Venn diagrams with radius 0.95 and centres 2.0 apart — the circles missed each other by 0.1 and never intersected. INNER JOIN, whose entire meaning is the intersection, rendered as two blank circles. It passed the compiler, passed the prose review, and shipped into a 113-page PDF before a human noticed.

Reading the .tex cannot catch that. Only looking at the picture can. So the skill renders every figure to its own PNG (skill/scripts/render-figures.py) and hands them to an agent that reads them as images and checks the geometry against the caption.

Style

Chapters are written for an engineer with a few years' experience — conversational, concrete, short paragraphs, code in small annotated chunks, no beginner hand-holding and no repetition. The LaTeX template ships O'Reilly-ish callout boxes (insight, practical, warning, aside) and listings syntax highlighting.

Customising

The whole skill is one Markdown file. After installing, edit ~/.claude/skills/youtube-to-book/SKILL.md — change the writing style, swap the callout colours, adjust the LaTeX preamble. Claude reads it fresh on every run.

Contributing

Issues and PRs welcome. To work on it locally:

git clone https://github.com/dev-shetty/youtube-to-book
cd youtube-to-book
node bin/cli.js install --force    # install your working copy

License

MIT © Deveesh Shetty