How to Write a Veo 3 / Google Flow Prompt That Goes Viral

Every viral AI clip starts life as a paragraph. Veo 3 and Google Flow follow direction astonishingly well — and expose its absence just as fast. When a generation comes out flat, the model is rarely the problem; the prompt is. This guide teaches the exact formula behind every master prompt in the KreoVid library, so you can write your own instead of gambling on one-line ideas.

Why most AI video prompts fall flat

Open any feed of AI video attempts and the failures share a shape: "a cute robot drinking coffee, cinematic, 4k, amazing." There is no moment, no change, nothing for the model to do — so it produces a screensaver, filling every gap with the most ordinary choice available. Viral clips are the opposite: something specific happens, on a schedule, and the last second gives the viewer a reason to replay or share. That structure can be written down. Here it is.

The master-prompt formula

A KreoVid master prompt is built from six blocks, always in the same order. Miss one and the clip wobbles; include all six and even an ordinary idea becomes reliably filmable.

1. Subject and scene

Name exactly who or what is on screen, where they are, and one visual signature that makes them memorable: the orange vest, the oversized hat, the impossibly tiny ladder. If the concept plays with scale — miniature worlds, giant props — anchor it against something normal-sized in the same sentence, so the model grasps the contrast.

2. Timed action beats

This is the block everyone skips and the one that matters most. Split the clip into timestamped beats: 0–3s is the hook — the most striking image of the whole video, front-loaded because viewers decide in the first seconds. 3–6s is the complication: something changes, fails or escalates. 6–10s is the payoff: the rescue, the reveal, the reaction shot. One clear action per beat — the model follows schedules far better than stories.

3. Camera

Give the camera a job: a slow push-in, a low-angle tracking shot, a locked-off frame the action plays inside. Unspecified cameras drift; specified cameras feel directed. For miniature scenes, a macro lens and shallow depth of field sell the scale instantly.

4. Lighting and style

One lighting cue (golden hour, neon rain, harsh flash) and one style anchor (photorealistic, tilt-shift diorama, claymation, VHS home video). Two short phrases here replace a paragraph of adjectives.

5. Sound

Veo 3 generates audio, so direct it: the ambience, the one sound effect that lands on your payoff beat, or deliberate near-silence before a sting. Undirected audio is random audio.

6. The negative prompt

Finish by banning what you keep getting by accident: no text, no watermark, no face changes, no extra fingers, no morphing. Negative prompts are cheap insurance and the difference between a usable clip and a near-miss.

Worked example: Tiny Workers, beat by beat

Our library prompt Tiny Workers — Bridge Rescue by Giant Hand is the formula in its purest form, and the most-copied pattern in the Miniature World collection. It opens by locking the format — vertical 9:16, ten seconds, tilt-shift miniature world — then introduces its subject with the scale anchor built in: construction workers two centimetres tall, in orange vests, building a wooden bridge across a garden puddle that reads to them as a canyon river.

The beats run exactly on formula. In 0–3s the crew works calmly — planks carried, tiny nails hammered, a mini dump truck rolling onto the half-finished bridge — with the hook already visible in frame one. In 3–6s a plank cracks and the truck tilts toward the water. In 6–10s comes the payoff the clip exists for: a giant human hand descends, lifts the truck to safety and holds the bridge steady while the crew cheers up at it. Camera (macro lens, slow dolly), lighting (golden hour) and a two-item negative prompt close it out.

Notice what you can steal: not the bridge, but the skeleton. Swap the crew for chefs or firefighters, the hand for a paw or a crane — the timing still carries the clip. Here it is as a fill-in template:

Vertical 9:16, [DURATION]-second video, [STYLE ANCHOR]. [SUBJECT with one visual signature] in [SETTING, with a normal-sized anchor if scale matters].
TIMING — 0-3s: [HOOK — the most striking moment, visible immediately]. 3-6s: [COMPLICATION — something cracks, fails, escalates or surprises]. 6-10s: [PAYOFF — rescue, reveal or reaction].
CAMERA: [shot type + one movement]. LIGHTING: [one lighting cue + mood]. SOUND: [ambience, plus the one effect that lands on the payoff].
NEGATIVE: no text, no watermark, no [the artefact you keep getting].

Common mistakes that quietly kill good prompts

  • Burying the hook. Your best moment sits at second eight, where almost nobody is still watching. Move it to frame one and let the rest of the clip explain it.
  • Ten actions in one beat. The model blends simultaneous instructions into mush. One verb per beat; escalate across beats instead.
  • No scale anchor. "Tiny workers" alone renders as normal workers on a normal set. A giant hand or a bottlecap — contrast is what the eye reads.
  • Forgetting the sound. On Veo 3, silence about audio does not mean silence in the clip. It means whatever the model felt like.
  • No negative prompt. Gibberish text, warped hands and drifting faces are the default failure modes. Ban them explicitly, every time.
  • Rewriting everything between retries. If a clip is 80% right, change one block and regenerate — shotgun edits teach you nothing about which line did the work.

Frequently asked questions

How long should a Veo 3 prompt be?

Long enough to cover the building blocks — subject, timed beats, camera, lighting, sound, negative prompt. In practice that is roughly 80 to 180 words for a 10-second clip. Shorter, and the model improvises the gaps; longer, and later instructions get ignored. When in doubt, cut adjectives before actions.

Should I use text-to-video or start from an image?

If the first frame's look matters — a specific character, a product, your own face — start from an image: generate or upload the opening frame, then let the prompt direct the motion. For simple, atmospheric scenes, text-to-video is faster and perfectly fine. Many KreoVid prompts note when a photo is needed; the rest work either way.

Why does the model ignore the end of my prompt?

Usually because the prompt is one long, undifferentiated paragraph: the model averages it and the final beats dissolve. Timed beats fix most of it — each line becomes a separate instruction with a moment attached. Also generate at the duration you wrote for; a 10-second beat plan squeezed into 5 seconds drops its own ending.

Start from a working skeleton

Writing from a blank page is the slow way. Take a finished master prompt, keep its structure and swap the subject, setting and payoff for your own idea — that is how most of this library was written. These are good skeletons to start from:

← Browse the full prompts library