Which thirty seconds of your video is worth posting?

Long video into short clips is the most useful thing AI does for small businesses with a camera, and the standard implementation makes one specific mistake: it looks for the interesting moment.

Interesting and selling are different moments, and usually minutes apart.

Why one score hides the decision

A single virality number ranks the funny bit first. The funny bit gets views, and the product is the setting rather than the subject, so nothing happens afterwards — and the business concludes short-form does not work for them.

So scoring here is eight separate commercial dimensions rather than one average. That lets you ask the question you actually have: which moment sells, which builds trust, which answers an objection, which works with no sound.

The clip you post for reach and the clip you put on a product page are different, and a blended score cannot tell you that.

The dimension that returns nothing

A dimension whose inputs were never measured returns null and says so.

This was a real change in behaviour. The engine previously produced a number for every dimension whatever it had — and a score computed from nothing is not conservative, it is a fabrication with a decimal point on it, and it is indistinguishable from a real one on the screen.

A visible gap is more useful than a confident guess, because a gap tells you which input to supply.

Moments from transcripts, not from trust

Candidate moments come from real transcript segments. Every recommendation therefore points at the sentences that produced it, and you can read them and disagree.

Before that, the ranking depended on a caller handing over timestamps — which nothing could actually produce, so the clever scoring sat behind a step that never happened. Worth saying plainly because it is this codebase's commonest defect: a finished component whose input had no producer.

What to record so this has something to work with

The limit is honest and it is the main one: it ranks what you recorded. If the source video has no sentence worth thirty seconds of a stranger's attention, no clipper will find one.

Three things that reliably produce clippable moments, and none need a studio:

• Answer one customer question per video, fully, as if the person is in the room.
• Say the price out loud. Nothing holds attention like a number nobody else will state.
• Describe a job that went wrong and what you did about it. That is the trust dimension, and almost nobody records it.

Then the dull mechanics

Reframe, captions, B-roll: model-gated and separate from the decision. Captions matter more than most people think — most viewing is silent, and a clip whose selling point is spoken and uncaptioned sells nothing.

The question to ask any clipping tool

Show me why this moment, not the next one.

If the answer is a single score, you cannot tell whether it picked the moment that sells or the moment that gets watched — and you will find out a month later from your order book.