Two Of These Eight Work On A Real Video. Both Hand You Notes.
“Delete CapCut. Forget Premiere Pro. It’s 2026.” Eight prompts for Google’s Omni model, posted in full in the comments: fix the lighting, blur anything sensitive, cut the dead air, find the five most viral clips, add captions, remove an object, fix the audio, make it vertical.
First, a correction to ourselves. A fortnight ago we looked at a similar card from the same page and pointed out that not one of its ten prompts cut, captioned, scored or exported anything — that they were all shot treatments wearing an editor’s name. This card fixes that. Number three cuts. Number five captions. Number seven is audio. Number eight is format. Somebody filled in the missing half.
Which raises a better question than the old one. Not does it edit, but how long a video can it edit — because Google documents two completely different answers.
The source post — “Delete CapCut. Forget Premiere Pro.”, BANI Academy. Eight Omni video prompts, each posted as its own comment.
Two machines, one card
There are two separate things you can do to a video with these models, documented on two separate pages, with two limits that are not close to each other.
Reading a video. From Google’s video understanding documentation: “Models with a 1M context window can process videos up to 3 hours long by default (at low media resolution), or up to 1 hour long at high media resolution.” You can refer to moments in MM:SS format, and ask for timestamps back.
Editing a video. From the Omni documentation: “Input videos for editing and extension must be 10 seconds or less when uploading.” Extension only appends to the end, to a total of forty seconds, and “all generated videos include SynthID watermarking.”
Three hours against ten seconds. That is a gap of about a thousand to one, and every prompt on this card sits on one side of it or the other.
Diagram: which machine each prompt is talking to
| Reading the video | Editing the video | |
|---|---|---|
| Documented limit | Up to 3 hours at low resolution, 1 hour at high | 10 seconds or less of input |
| What you get back | Text — timestamps, summaries, selections | A re-rendered video file |
| Can you check it? | Yes. Scrub to the timestamp and look | Only by watching, frame by frame |
| Marked as AI? | Nothing to mark — it is notes | Yes — SynthID on every generated video |
| What it costs | Tokens, roughly 100 per second of video at low resolution | A generation, plus the wait |
| Prompts on this card | 2 — find the viral clips, decide what to cut | 6 — lighting, blur, captions, removal, audio, vertical |
All eight, against a twenty-minute video
Take a real file — a twenty-minute talking-head video, the thing anybody writing “delete CapCut” has in mind — and run the card down it.
| Prompt | Which machine | On a 20-minute video | What it hands back |
|---|---|---|---|
| Find the most viral clips five ~30s moments, exact timestamps, hooks, headlines |
Reading | Works. Well inside the three-hour ceiling | Five timestamp pairs and a headline each. A shot list |
| Cut the dead air remove pauses, filler, false starts; report both runtimes |
Reading to decide, editing to render | Half works. It can tell you what to cut; rendering the cut is the ten-second path | An edit decision list, if you ask for it that way |
| Fix poor lighting | Editing | Ten seconds of it | A graded clip |
| Blur sensitive details “review of every frame” |
Editing | Ten seconds of “every frame” | A blurred clip |
| Add captions | Editing | Ten seconds | A captioned clip, burned in |
| Remove anything from the shot | Editing | Ten seconds | A cleaned clip |
| Fix bad audio | Editing | Ten seconds — and audio repair is a talk-length job | Cleaned audio |
| Make it vertical | Editing | Ten seconds — and reframing is a whole-video job | A 9:16 clip |
Notice what that does to the headline. The one prompt that survives a real video does not produce video. It produces five pairs of timestamps, a hook for each and a suggested on-screen headline — which is a brief for an editor. You read it, open CapCut, and cut to those marks.
So “delete CapCut” is exactly backwards. The working half of this card is a machine that watches your footage so you do not have to, and writes down where the good bits are. That is genuinely valuable — scrubbing a twenty-minute recording for clips is the dullest hour in the week — and it ends with you opening an editor.
The two best-written prompts are the two that check themselves
This is worth saying plainly because it is unusual. The two reading prompts both contain an instruction that lets you catch the model being wrong.
“Find the most viral clips” ends with: “Never invent content that isn’t in the video.” That is precisely the failure mode of asking a model to report on a long recording — a confident quotation of something nobody said, at a timestamp where something else is happening. And because the output is timestamps, you can check every one in thirty seconds by scrubbing to it.
“Cut the dead air” ends with: “Report the original duration and final runtime.” Two numbers you can compare against the file. If it claims it trimmed a 20:14 video to 17:02, that is a checkable claim, unlike “I removed the filler words”.
Both are the same technique, and it is the one worth stealing from this card: ask for an output you can verify without re-watching the whole thing.
Two prompts here are not video jobs at all
Look again at numbers three and four and notice what they actually operate on. Removing filler words, false starts and repeated phrases is an operation on speech. Finding the five strongest moments is an operation on what was said. Neither needs a model to look at pixels.
Which means you can do both today, at any length, with a transcript and any capable text model: no ten-second ceiling, no render queue, no watermark, no upload of the footage itself. Feed it a timestamped transcript, ask for the clip boundaries and the cut list, and you get the same deliverable with none of the constraints — and it is cheaper, because you are paying for words rather than roughly a hundred tokens per second of video.
The workflow the card is actually describing
Read it as a division of labour rather than a replacement and it becomes a good week’s system.
| Stage | Send it where | Why |
|---|---|---|
| 1. Watch the footage | The reading path — or a transcript and a text model | Three-hour ceiling, output is timestamps you can verify, costs nothing to redo |
| 2. Decide the cuts | Same place, in the same pass | Ask for an edit decision list: in point, out point, one line of why |
| 3. Make the cuts, captions, audio | Your editor — CapCut, Resolve, Premiere | No length limit, no re-render of the whole file, and you keep the original |
| 4. The one hero shot that needs rescuing | The editing path, ten seconds at a time | This is what the ten-second limit is actually good for: the clip you could not fix before |
| 5. Export | Your editor | Nothing on this card exports anything |
And on stage four it is worth remembering what comes back: “all generated videos include SynthID watermarking, which is invisible to viewers but can be detected programmatically.” For a rescued shot inside a longer video that is a fact to know rather than a problem — but it is a fact, and it travels with the file.
Four honest caveats
- We did not run the eight prompts. This is a reading of what each one asks for against Google’s two documented paths and their two documented ceilings. We did not upload a video to either.
- The split is ours, not the card’s. Nothing on the card says “this one is analysis and that one is editing”. We assigned each prompt by what it asks the model to produce — text or pixels — and you may disagree on prompt three, which straddles both.
- Limits move, and these two pages move independently. Both figures were read in October 2026: three hours for understanding, ten seconds for editing. The editing ceiling in particular is the sort of number that gets raised, and if it does, most of this article’s arithmetic changes. The two-machines point does not.
- “Google Omni” is the card’s name for it. Google’s own pages use Omni and Omni Flash for the generation and editing model, with video understanding documented separately. If you go looking, search for both.
The one-line version
Eight prompts, two machines. The one that reads your video handles three hours and hands you timestamps; the one that edits your video takes ten seconds and hands you a watermarked clip. On any real recording, the prompts that work give you notes, and the prompts that give you video do not reach the end of the first sentence you said.
Which makes this a very good card read the right way round: let it watch, and do your own cutting. That is also the honest version of the claim we checked a fortnight ago — and this deck, unlike that one, at least asks for the cut.
Try the one that works. Take your last long video, run prompt four, and check two of the five timestamps by scrubbing to them. If both land where it said, you have just replaced the worst hour of your editing week — and you still open CapCut afterwards.
