Gemini Omni turns a written prompt into a short video clip, and the clip is only as specific as the prompt you wrote. The prompts behind the demos that went viral share one shape: subject, motion, camera, light, duration, then constraints. This page teaches that shape, the failures that waste your generations and the settings I check before every run; the free download adds 18 templates and a mini-course.
What Gemini Omni makes
Omni makes one clip per prompt. You describe a scene, it renders a few seconds of motion, and you either keep the result or change the prompt and go again.
The clips that get watched fall into five buckets. Product reveals, where a single object turns or opens in good light. Character moments, one person doing one thing with a visible emotion. Establishing shots that set a place before a story starts. Loops for ads, where the last frame matches the first. And camera moves a real crew could never afford.
The prompt structure behind the viral demos
Every good Omni prompt I have collected answers six questions in the same order. I write them as labelled lines, then keep the labels or fold them into a paragraph. The labels are mine; Omni reads plain language.
Subject: who or what, with two physical details
Motion: one action, with a start and an end
Camera: where it sits and one movement, or "static"
Light: the source, its direction, the time of day, the mood
Duration: how many seconds, and what the last frame shows
Constraints: aspect ratio, visual style, what to leave out
Subject comes first because it is the first thing I check. "A woman in a yellow raincoat, hood up" gives the model something to hold. "A person outside" gives it nothing.
Motion is one action. One. If the subject needs to pick up a cup and then walk away, that is two clips. Every time I stack a second action into a prompt, the subject morphs somewhere in the middle.
Camera is the line people skip, and it separates a demo from a home video. Name the position (eye level, low, overhead, over the shoulder) and one movement (static, slow push-in, orbit, handheld follow). If you say nothing, the model decides, and it decides differently every run.
Light does more for realism than any style word. "Golden hour, backlit, long shadows, warm" reads as film. "Overcast and soft, no shadows" reads as catalogue. Saying nothing gets you neither.
Duration and the end frame belong together. A clip that ends mid-gesture feels broken, and a clip that ends on a held frame feels finished. Write what the final frame shows.
Constraints go last: the aspect ratio for the platform and the visual style in two words, then the things to leave out. I add "no text, no logo" almost every time, because lettering is where video models fall apart.
The same six fields, folded into one prompt:
A matte black ceramic mug on a wooden counter, steam rising. The mug stays
still while the steam curls upward and a hand enters from the right to lift
it. Camera: static, eye level, shallow depth of field. Light: warm morning
window light from the left. Duration: about six seconds, ending as the mug
leaves the frame. Constraints: 9:16 vertical, photoreal, no text, no logo.
Read it aloud. If it takes more than one breath, cut a detail. Short prompts with all six fields beat long prompts with four.
Which settings matter before you press generate
The settings panel changes between versions, so look at yours before trusting anything I list here. In any video tool I use, four settings decide whether a clip is usable before the prompt runs: aspect ratio, matched to where the clip will live (vertical for Reels and Shorts, wide for YouTube and decks); duration, set to the shortest length that fits one action; a reference image, if the tool accepts one, because a still you already like does more for consistency than a paragraph of description; and resolution, kept low while you iterate and raised only for the final render.
Change one setting or one prompt line per run. Change four things at once and the result teaches you nothing. When a clip is right, save the prompt next to it; next week you will have forgotten the wording.
Common failures and the fix for each
| What you see |
Why it happens |
The fix |
| The subject changes shape halfway through |
Two actions in one prompt |
One action per clip; make the second action a second clip |
| The camera drifts or cuts on its own |
No camera line |
Name the position and one movement, or write "static" |
| Flat, grey, lifeless footage |
No light line |
Name the source and the time of day |
| Lettering on signs or packaging comes out garbled |
Video models struggle with text |
Leave text out of the prompt and add it in an editor afterwards |
| The clip stops mid-gesture |
No end frame named |
Write what the last frame shows |
| A prompt that worked yesterday gives something different today |
Generation is random by design |
Keep the prompt, change nothing, run it two more times before you edit |
What is inside the Gemini Omni guide
The free download has four parts.
- Gemini Omni Prompt Engineering Guide: the six-field structure above, line by line.
- 18 Prompt Templates plus Viral Hit Prompts: fill-in templates for the clip types that get shared.
- Gemini Omni Prompt Engineering Mini-Course: short lessons from a first clip to a repeatable session.
- Lifetime access and updates: every revision as the model changes.
Drop your email in the form and it is yours.
For the tools around Omni: the Gemini guide covers the assistant itself, the Veo guide covers the same fields written as JSON, the Nano Banana guide covers stills and edits, and the Gemini prompts in the library cover the text side.
