Turning a long video into a short one is a selection problem

A blunt tool would carve the two hours into 180 boxes and let you pick one. Seven thousand two hundred seconds divided into forty-second pieces is a hundred and eighty of them, and almost any long-to-short conversion is really that choice wearing a different interface. This one does not slice at all: the transcript is split at the speaker’s own sentence ends and pauses, and a clip can open or close anywhere between two words — never inside one — as long as a pause or a sentence actually put a boundary there.

The whole recording is read, in parallel, not sampled. The transcript is split into windows of roughly thirty minutes, each read independently, with three minutes of overlap bled onto every inner edge — so a moment that straddles where one window ends and the next begins is whole inside at least one of them. A single pass over a long transcript anchors on whatever it reads first and stops picking well before the back half; the windows exist because that is measurable, not assumed.

The opening minutes are excluded, not guessed at. A model reads the start of the transcript and reports where the teaser, the channel branding, the host’s welcome or the guest pleasantries actually end, rather than a fixed sixty seconds being skipped and hoped to be enough.

What does not survive was read, scored, and turned down. Every candidate that gets this far carries four scores — hook, flow, value and Travel — its exact timecodes, and the sentences it was built from. The seconds that end up back in the two hours were not skipped; they were the ones a fresh read still ranked lower.

• A credit buys a minute of the recording read, not a clip exported — the meter is on analysis, and the transcript is cached against the source, so asking the same video for a different angle later costs nothing again
• Every plan includes a generous run of clips before an extra one draws on that same balance — about twenty per hour of the plan’s monthly allowance, roughly double what a typical hour of source actually yields
• Two clips can share part of the same minute — different points can live close together — but past half of the shorter one they are treated as the same clip twice, not published twice
• A cut that comes back outside twelve to a hundred seconds is made again from scratch, never trimmed to fit — length is a consequence of where the idea ends

One stage points, a second one cuts. The first pass over each window only says where a moment lands and what it is — no boundaries at all, so a real moment is never lost for want of a tidy range around it. A second stage, reading three minutes either side, decides the actual cut. Asking one pass to do both is where a worse version of this used to go wrong.

A boundary lands between two words, never inside one. The editor names the exact words it wants to open and close on, quoted verbatim from the transcript, and the timestamps are read off those words rather than estimated — the one thing "anywhere" must not mean, once there is audio attached to it.

Up to three pieces, in the order spoken. A setup from a minute earlier can be joined to the moment itself, with whatever sits between them left out of the clip entirely — used only when the jump a viewer will hear is worth what it buys, never to look clever.

A stranger reads it before you do. The finished lines alone — no transcript, no title, no timecodes — go to a cold read. A cut that cannot be followed on its own is sent back and cut again; one that still fails keeps its place in the list, with its score capped and a line saying exactly what a viewer would be missing.

Twelve to a hundred seconds, refused rather than sheared. The editor aims for something in the middle of that band, and a cut that comes back outside it entirely is made again, not trimmed down — the length is an outcome of where the idea finishes.