What the tools promise
The logic of clip-cutting tools is always the same: you upload the long video, the system transcribes it, finds the "interesting" parts in it, cuts them out, reframes the picture to vertical, burns subtitles on it, and hands back ten to twenty finished short videos. Some of them even attach a score for which one they think will perform.
Most steps in that chain genuinely work well. The transcription, the cutting, the vertical crop, the speaker tracking and the subtitles are all reliable, mechanical work. What does not work well is precisely the one step you bought the whole thing for: the selection.
The one independent number that exists
There is no independent quality measurement in this field — every "benchmark" that surfaces comes from a company that sells a clip cutter. There is a single exception: a 2026 academic study that compared human annotations with machine suggestions.
- The machine's best guess matched the humans' roughly 26 percent of the time.
- The humans themselves agreed with each other only to a degree of between 0.22 and 0.72.
- So it is not only the machine that is uncertain: the task itself is badly defined.
The second point is the more important lesson. If two people cannot agree on what the best thirty seconds of a video is, then the machine cannot be expected to "get it right" either. What you can expect from it: that it puts twenty candidates in front of you, and you choose. That is genuinely useful — it is just not the same thing the marketing promises.
Why the machine does not know what is interesting
Because the machine works from the transcript, and what it measures in the text is what sounds "quotable": a self-contained thought, a rounded sentence, a strong claim. That is not a bad heuristic — but it is not what the success of a short video turns on.
A short video turns on whether there is a reason to keep watching in the first second. That can be a surprising number, a contradiction, an unspoken question — something the text does not show, because it comes from the context. The machine does not know that a given sentence lands because three minutes earlier somebody claimed the opposite. You do.
The recipe
1. TRANSCRIPT FIRST, and review it. The cutter picks from the
transcript — if the transcript is bad, so is the pick.
2. ASK FOR MANY CANDIDATES, not few. Four good ones out of
twenty is a normal ratio. So is two out of ten.
3. YOU CHOOSE what goes out. This step cannot be delegated.
The question is not "is this a good sentence", but
"would this stop somebody mid-scroll".
4. RECUT THE OPENING. The machine starts at the top of a
sentence; you have to start in the MIDDLE, where the tension is.
5. SUBTITLES ARE MANDATORY, and review them. It is watched muted.
Fix the names and the numbers here.
6. LABEL IT if part of it was made with AI.
The traps of the vertical crop
- Several speakers. The automatic tracking will jump between two people. If two of them are talking, set a split screen or a fixed crop instead.
- Text on screen. Slides, a whiteboard, a screen recording — these fall outside the vertical crop and become unreadable.
- Gestures. If the speaker points at something, a tight crop cuts off the hand, and the sentence stops making sense.
- Where the head sits. The machine centres it; on a good short video the head is higher, so the subtitle has room underneath.
What to put at the front of the clip
The first second of the clip decides, and this is where the machine does not help. What people do, and what works: the clip starts in the middle of the answer, not at the question. Not "let's talk about how transcription works", but "every fifth word will be wrong — and the vendor does not put that on the page".
The second trick is to supply the context with a caption, not with speech. One line at the top of the frame says what this is about, and it saves you the six seconds of preamble that make the viewer scroll on. The general logic of breaking a longer piece into several formats is what One Piece of Content into Ten Formats lays out in full.
Common misconceptions
- "The machine finds the best parts." By the one independent measurement, once in four. It gives candidates, not decisions.
- "If I make lots of clips, one of them is bound to hit." Volume on its own does not bring attention — the quality of the opening does.
- "The machine does the subtitles, done." It gets the names and the numbers wrong, and it is watched muted. Reviewing is mandatory.
- "The vertical crop is automatic." With one speaker, yes. With two, with a whiteboard, with a gesture, it needs a hand.
What to take away
A clip-cutting tool does the mechanical work well — transcript, cut, crop, subtitles — and does badly the one thing you bought it for: choosing. Ask for many candidates, choose yourself, and recut the opening. On the quality of the transcript, which the whole chain rests on, Transcription and Subtitles is a post of its own; and for the labelling duties there is Labelling AI Video, which walks through the rules.