YouTube lecture to notes: summary first, transcript with timestamps.

The file opens with a YAML block holding the video title, the channel as author and the video's URL as the source, so the note always leads back to the video. Then come three sections. Summary is three to five sentences on what the lecture covers. Key takeaways is a short list of the points worth remembering, between three and seven. Transcript is the full lecture, where every paragraph starts with the time it was said, such as [12:40], or [1:05:12] once the video passes an hour.

For lectures longer than about a quarter of an hour, or with clear changes of topic, the transcript is split into chapters with their own headings, so the outline of the file reads like a table of contents. When more than one person speaks, the paragraphs are labelled Speaker 1, Speaker 2, or by name if the speakers introduce themselves. Passages that can't be made out are marked [unintelligible] instead of being guessed.

We don't download captions. The link is passed to Google's Gemini, which watches and listens to the video itself, so it works on lectures that have no subtitles, and the automatic ones YouTube generates don't leak their mistakes into your notes. The transcript stays in the language of the lecture and isn't translated. A video with no speech at all, such as a silent demo, gets a description of what happens on screen and a warning saying so.

The timestamps are plain text. They don't link back to the video. To jump to a moment, add &t= and the time in seconds to the video URL, or scrub to it by hand.

YouTube conversion is part of Premium, which costs $15 a month or $12 a month billed yearly. A Premium account gets 10 audio, video or YouTube conversions and 120 minutes of media per day, and the counters reset at midnight UTC. The video has to be public. Private, unlisted and age-restricted videos can't be processed, and the practical limit is around 45 minutes. Processing stops after 15 minutes, so a very long video may fail rather than finish, and a shorter cut of it works more reliably.

Long transcripts sometimes run into the model's length limit. When that happens the request is retried once from the start, and you never receive half a transcript presented as a whole one. Links to Shorts and live recordings work. In a playlist link only the video in the v= parameter is converted, not the whole playlist.

If you have the lecture as an audio file instead, the result has a summary and a transcript, split into topics when the recording has clearly distinct ones, but no timestamps.