Summarise long videos from their transcripts

Turns a stack of hour-long videos into a searchable table of summaries. You list the transcripts you want read, and each one is fetched, summarised with the key points pulled out, and added to a Google Sheet as its own row alongside the date, the link and how much of the transcript was covered.

A whole transcript rarely fits in one model call, so it is cut into sections along its own line breaks; each section is read on its own and the notes from every section are then written up as one summary. That is what keeps the ending of a two-hour talk in the summary instead of only its first ten minutes. 'maxSectionsPerVideo' caps how far a single video can go, so one enormous transcript cannot run up an unbounded bill; when it bites, the run says so and the row records '10 of 14' rather than pretending the whole thing was read.

Lodol cannot pull a transcript out of a video, so this workflow reads transcripts you point it at: any address that returns the text will do - a .txt, .vtt or .srt caption file, a published transcript page, or a file you have saved somewhere it can be fetched from. Write one per line in the 'transcriptLinks' step as 'Video title | https://link-to-the-transcript'; a line that is only a link works too and is labelled with the link. A transcript that cannot be fetched, or comes back empty, is reported and skipped, and the rest of the list still goes through.

Set it up by filling in 'transcriptLinks' and 'reportSheetUrl'. Two things are worth tuning: 'summaryStyle' decides what the summary reads like - change it to ask for a lecture recap, a research note or show notes - and 'charactersPerSection' decides how much transcript goes into each model call. The model is set to GPT-5.4 mini because summarising is cheap work at volume; switch it on the step if you want a more capable one.