One recording, nine posts: the Offcut pipeline end to end

I record one forty-minute session a week. This follows a single recording all the way through: transcript, candidate cuts, hook ranking, vertical reframe, and the approval list where I say yes or no in about ninety seconds.

I show the rejects too, because they are the interesting half. Two in three candidate cuts are bad, and the reason is consistent: the model over-selects for a clean sentence, and the cleanest sentence is very often the least interesting thing you said. The ranker now penalises fluency. That one change roughly doubled the keep rate.

All of it runs on a laptop and two APIs. No proprietary model, no fine-tune, no moat. If you want to build this yourself, this video is enough — and if you would rather not, that is what Offcut is for.