linzhuochen/video-to-notebook

video-to-notebook is a Python command-line tool that turns a few playlists of open-course videos on the same topic into a single static website that reads like a merged textbook with an attached concept encyclopedia. The author points it at YouTube or Bilibili playlists, then a coding agent such as Claude Code, OpenAI Codex, Cursor, or Continue drives the pipeline: it crawls the videos, tags every transcript chunk against an ontology, clusters those tags into a concept graph, plans a chapter order with teaching in mind, and writes each chapter and concept page. Output ships as an Astro-based static site with Pagefind search and optional bilingual Chinese and English versions. A distinctive design choice is that no separate Anthropic API key is required by default. The CLI writes a prompt envelope for each stage to a state directory and exits, and the agent then writes its decisions to a sibling file and re-runs the command with an apply flag. The README also offers an Option B that uses an Anthropic API key directly. Required tooling is Python 3.12 or newer, and the build uses Astro 5 for the static site. Three design principles run through the README. First, source fidelity: each chapter and concept page must reflect how the lecturer actually explained the idea, including verbatim phrasings, analogies, and named citations, and any framing the agent adds on top must carry an explicit marker. Second, no fabrication: when source chunks are too thin, the agent is told to stop and debug the pipeline rather than fill the gap from training data, and common bug patterns are listed for the agent to check. Third, textbook-note depth rather than magazine-summary depth, with target chapter lengths, numbered sections, math derivations annotated line by line, colour-coded callouts, a runnable PyTorch skeleton when a model is introduced, and five to seven takeaways anchored to lecturer examples. The rendered output includes a textbook table of contents, a chapter reading view with a left sidebar, a TL;DR callout, a right-side mini-map, and deep links back to the source video clips, alongside an encyclopedia of illustrated concept pages with interactive widgets and counter-example misconceptions. The project is MIT licensed and is in early development.