Long form to short form is mostly choosing which moments to clip
Software cuts, reframes and captions short clips from a long video by itself these days. The hard part is picking the moments: on a three-day event shoot we worked on, a per-clip score couldn't get 531 candidates down to the 25 the brief asked for.
Deploy OpenVideo
Long form to short form, and the part people still pay for
Long form to short form is the job of cutting one long recording into a set of vertical clips for YouTube Shorts, Instagram Reels and TikTok. The recording might be a 45-minute podcast episode, a webinar, a keynote or a whole day of event footage. Some people call it clipping, or cutting a long video into shorts, and agencies and production teams tend to call the results cutdowns. On YouTube, anything square or vertical and up to three minutes long counts as a Short, as long as it was uploaded on or after 15 October 2024.
Businesses do it for reach. One founder started filming his podcast and posting clips of it on every short-form platform. A year later he counted 1.2 million followers across platforms, against 55,000 a year earlier on his social account and newsletter combined.
The editing has mostly been automated. Software finds candidate moments in the transcript, cuts on sentence edges, reframes a landscape shot to 9:16 around whoever's talking and burns in captions. One podcaster who records with remote guests said his edit fell from about two and a half hours to 25 minutes once the multicam cut and the clips were done for him.
The brief is another matter. Look at what people hiring for this work ask for:
- a retailer's founder records 4 to 6 minute talking-head videos and wants 3 to 5 clips from each
- a social media marketer wants 10 to 15 clips per podcast episode
- one advertised role covers a long YouTube video every week and 90 short clips a month
- a baseball creator with about 275,000 subscribers pays $100 for a long edit and $20 for each short one
Every one of those asks for a number of moments out of a longer whole. Picking them is what people still hire a person for. A recruiter filling one of these roles wrote that the client wanted an editor fluent enough in English "to understand what clips to create", and would turn away anyone who used AI for the cuts or the captions. The host of a long-running interview podcast advertised a full-time job just making clips, and wanted an editor "who knows how to hook people".
531 candidate moments for 25 slots
We got a close look at this in October. An agency was filming a three-day conference for its client, and its editors wanted the best moments of the shoot proposed to them so they could cut them in their own software. The footage came in the way event footage does: a folder per camera, cinema-camera originals in 4K, testimonials tucked one folder deeper. Day one alone was 275 video files across 12 folders. By the end of day three there were about 1,100 clips.
We built the selection step for them into OpenVideo, our open-source footage editor. A model watched each clip once and wrote a log: what's in the shot, whether it's an interview, a stage shot or b-roll, what's wrong with it, and the best quotes with their times. On day one's first pass, 266 clips were logged in about two and a half hours. Then a second pass read the logs and transcripts against the agency's brief and proposed moments, either a soundbite with its words and speaker or a piece of b-roll with the shot described. Each one got a score from 1 to 5 and a reason.
It proposed a lot. Day one produced 296 candidate moments, day two 477 and day three 531, and day three's brief ended with "Max 25 clips". Of day three's 531 candidates, 196 scored 4 or 5 and 19 scored 5, so the scores could put the list in order but could never cut it to 25. Nor could they tell when two cameras had caught the same answer, since the reading never put clips from different camera folders next to each other.
One more pass got it to 25. It read every remaining candidate side by side against the brief and ranked them: the best 25, then stand-ins, then everything else. That took about half a minute and, at the model's listed price, two to three cents. What came back was 12 soundbites and 13 pieces of b-roll from 22 different clips in five folders, never more than three from one clip. Where two cameras had filmed the same interview, each camera supplied different lines.
A virality score answers a different question
Many clipping tools handle the choosing with a virality score. Every clip gets a number, often out of 10 or 100, and some tools will post anything above a threshold you pick without checking with you first. The number is supposed to tell you which clip will travel.
The research doesn't back that much confidence. In a 2026 study, researchers had two multimodal models, Gemini 3 Flash and Qwen 3 Omni, rate short videos the way 673 particular viewers would, on a scale covering emotional arousal, dramatic impact and novelty. Then they checked each rating against what that viewer had actually said.
| Who rated the video | Agreement with the viewer |
|---|---|
| Either model, no examples given | 0.14 |
| Gemini, shown examples first | 0.23 |
| Another person with a matching profile | 0.29 |
Agreement is an intraclass correlation, where 1 would mean identical ratings. The figures come from Multimodal Large Language Models as Synthetic Participants in Video-Based Studies.
The models land well below a stranger. And the stranger isn't much of a predictor either: two similar people agreed at only 0.29 on how much the same clip grabbed them. A single number per clip papers over that disagreement.
An agency answers to a brief: this speaker, this product, this message, nothing said twice. A model can check a transcript against that, and a person can check the model in seconds. On our shoot the 1 to 5 score was good for sorting. We'd keep it to that, and we'd never let a score publish a clip by itself.
How we got from 531 to 25
These are the steps that ran on the shoot. Each clip was watched once, to log it, and everything after that reads text, the logs and the transcripts. That's what kept the choosing cheap: watching every clip again would have cost as much as the logging did, and reading the text cost cents. How the watching works is covered in our piece on AI video analysis.
| # | What happens | Where it stops |
|---|---|---|
| 1 | A model watches each clip once and logs it | Files over 30 GB were refused by the video host |
| 2 | It reads the logs and transcripts against the brief, a folder at a time | Long interviews come back thin without a density hint |
| 3 | Each soundbite snaps to whole sentences, plus a breath | A few still start mid-thought |
| 4 | One pass orders every candidate for the brief | It listed its picks in clip order, so sort by score |
| 5 | A person plays each pick and keeps or drops it | Review copies are 1080p, not the deliverable |
| 6 | The kept picks go to the editor as a sequence | No camera timecode in the file yet |
Step 2 reads a folder's clips together, twelve at a time, since the better of two similar shots, or one moment seen from another camera, can only be judged side by side. The density hint turned out to matter. A 45-minute interview gave six soundbites on the first reading. Told to expect about one every two to three minutes of talk, it gave ten, and one of them was the line the logger had already flagged.
Step 3 is the sort of detail editors notice straight away. Each soundbite starts 0.15 seconds before its first word and ends 0.25 seconds after its last, so no word gets clipped. In step 5 a pick plays from its in point, one key keeps it and another drops it, and dropping one of the best 25 pulls in the next without another model call. A fresh reading gets the kept and dropped picks back as examples of taste, and when those disagree with the brief, the brief wins.
One recording or a whole shoot: what to hand back
Long form to short form is really two jobs, and they need different output.
| One long recording | A multi-camera shoot | |
|---|---|---|
| Source | A podcast, webinar or talk | Hundreds of files, a folder per camera |
| A good clip | Stands alone, opens on its point | Fits the brief, no moment twice |
| Hand back | Finished 9:16 clips with captions | Selects for the editor's timeline |
| Finished by | Whoever posts, after a trim | The editor, from camera originals |
With one recording, finished vertical clips are what you want back. That's how OpenClipper, our open-source clipper, works. You give it a brief and a number of clips; left alone, it aims for about one clip per three minutes of video, close to the density that worked on the interview above. It cuts on the edges of speech, crops each shot around the speaker or keeps a screen recording whole, and burns in captions. Run it again and it skips the stretches already clipped or dropped.
A shoot is different. The review copies a browser could stream were 1080p at about 3.9 Mbps. The camera originals were 4K at about 62 Mbps, which puts a 70 to 80 minute take at 32 to 38 GB on its own. So anything rendered from the review copies is review quality, and the agency's editors wanted the picks in their own software anyway. The export is an XML sequence in the format Premiere Pro and DaVinci Resolve both import, with the soundbites and the b-roll laid end to end and each pick's words, score and reason attached as markers. The editor opens it against the camera originals and cuts from there.