yt2textbook: what one agentic model ships when the brief is a real open-source release.

The brief was deliberately commercial-grade: take any YouTube video or playlist and print it as a complete, illustrated textbook, the kind of document that replaces watching rather than summarizing it. Screenshots where the video shows something, learning objectives up front, comprehension checks and a glossary at the back, timestamped links to every source second, and a Word export. Then package the tool as an open-source release a stranger can run: README, examples, license, notices, and repository metadata that search engines can find.

Summarizer apps were deliberately not the bar. A paragraph about a 52-minute setup guide is useless when you actually have to set the thing up. The hard bar is a document you can work from, and that is what made the build worth measuring: the model had to carry ingestion, vision filtering, course planning, prose, and packaging, not just text in, text out.

The opening brief was sent at 7:53 in the morning. About 40 minutes later the pipeline ran end to end: transcript fetched, frames extracted and filtered, a course plan written, sections printed, and the first validated document on disk. By mid-morning the full example corpus existed, eleven documents from the Hermes Atlas training channel processed in parallel batches, with one module re-printed after the planner fix described below. By early afternoon the release was packaged: README restructured to the owner's direction, three diagrams drawn, a terminal demo recorded, license and notices written, repository created and described for search.

The human contribution was briefs, review gates, and four judgment calls. This is the working pattern the experiment set out to probe: a model left alone with a folder, a toolchain, and a definition of done, inside a loop where a person reviews every output before it ships.

• Working pipeline about 40 minutes after the opening brief
• Full packaged release about 4.5 hours later, corpus and marketing layer included
• A human review gate on every output

The measured output is about 1,050 lines of stdlib-first Python in a single module: transcript ingestion with a local Whisper fallback, frame extraction by scene-change detection plus interval sampling, perceptual-hash deduplication, a vision filter that caps classification at 24 informative frames per video, a course planner with a retry-and-fallback guard, section writers, and an assembly pass that adds comprehension checks, a glossary, and footnotes. The Word engine renders each document as a styled textbook with a cover page, a table of contents, real tables, and captioned figures. A deterministic normalization pass enforces the house style on every generated document, because the writer model would not.

The evidence pack is the corpus: eleven printed course documents from the Hermes Atlas training channel, reproduced with the owner's permission, holding 258 embedded screenshots, each captioned and linked to its second in the source video, plus eleven Word exports. The package around it is the release layer: README with three diagrams and a recorded terminal demo, Apache-2.0 with NOTICE, brand and security policies, third-party provenance notes, and repository metadata written for search.

Four material corrections were logged in the window, and none of them were code. First, the writer model kept emitting em and en dashes despite explicit prompt instruction; prompting did not hold, so a deterministic normalization pass in code did. Second, one masterclass module printed thin, about a third the depth of its peers, because the course planner answered prose instead of structured JSON; output review caught it, a retry-and-fallback guard went in, and the module was re-printed. Third, the product identity was a human call: the model's first name was not discoverable, and its first README buried the usage path, so the owner rejected both and the model rebuilt them to direction, name, structure, and search positioning included. Fourth, rights and privacy stayed human: choosing owner-permissioned source material, writing the provenance notes, and running the secrets review before publication.

Every one of those corrections is a judgment about context the model did not hold: what a name is worth, what a stranger needs first, what may legally enter a public repository. That boundary, not code volume, is what this experiment was designed to locate.

A printed page reads authoritative whether or not it is. The documents are model-generated from real transcripts and real frames, but the prose is not independently fact-checked, frame OCR can misread a blurry terminal, and videos without captions fall back to local transcription at higher cost. The corpus is also one channel's content: English, terminal-heavy, single-instructor. Slideshow talks, panels, and non-English sources are untested. The tool and this report both name that boundary, because treating a legible document as an authoritative one is exactly the mistake this program exists to avoid.

zorost/yt2textbook on GitHub→