How to write Veo prompts that work in Google Flow

What Flow runs today

As of October 2026 Flow offers four video models: Veo 3.1 Lite, Veo 3.1 Fast, Veo 3.1 Quality and Gemini Omni Flash 1.1. Google's developer documentation now treats Omni Flash as the default for most video generation and keeps Veo 3.1 for specific needs such as scene extension or last-frame control. The prompt advice below works for all of them.

Credit cost per clip in Flow (October 2026): Veo 3.1 Lite 10, Fast 20, Quality 100; Omni Flash 720p from 7 to 15 depending on length. Ultra subscribers pay half for Lite and Fast. Google says costs can change at any time, so check Settings in the prompt box before a big batch.

The structure Google recommends

Google's prompt guides list the same building blocks. You do not need all of them in every prompt, but each one you leave out is a decision the model makes for you.

Google Cloud's Veo 3.1 guide puts them in this order: camera work, subject, action, context, then style and ambiance. Clear, direct sentences work better than long, flowery description.

[Type of video]. [Shot size and camera move] of [subject with 2-3 details] [one action] in [place, time of day]. [Light and mood]. [Style].
Audio: [sound effects]. [Background sound]. [Speaker] says: [one short line].

Example prompts you can adapt

These are our own examples, written to the structure above. Swap the subject and the place and keep the shape. Set the aspect ratio (16:9 for YouTube, 9:16 for Shorts) and the clip length in Flow's settings, not in the text.

Realistic commercial clip. Slow orbit close-up around a matte black wireless earbud case on a wet slate surface at dawn. Soft side light, water drops catching the light, shallow depth of field. Clean, premium look.
Audio: quiet morning birds, a soft click as the case opens. No music.

Realistic, handheld medium close-up of a tired night-shift nurse in blue scrubs leaning on a hospital vending machine, a fluorescent tube flickering above. She looks into the camera, sighs and says: Twelve hours, and this machine still eats my coins.
Audio: the hum of the vending machine, a distant pager beeping.

Nature documentary style. Wide aerial shot slowly descending over a misty pine forest at sunrise, a river bending through the valley below. Golden light breaking through the fog.
Audio: wind in the trees, distant birdsong. No dialogue.

2D flat animation with bright pastel colors on a plain cream background. A cartoon coin with small legs climbs a staircase made of bar-chart columns, each step taller than the last, while the camera tracks sideways with it.
Audio: a playful tap on every step, a small cheer at the top.

Realistic overhead shot of two hands tearing a steaming loaf of sourdough bread on a wooden table, crumbs falling. Warm window light from the left, slow motion.
Audio: the crust crackling, quiet kitchen room tone.

Cinematic horror look, desaturated and grainy. Low-angle shot down a dark apartment hallway; at the end, a door slowly creaks open on its own while a bare bulb sways.
Audio: a long door creak, a faint water drip, then silence.

Tracking shot from knee height following a skateboarder from behind as she rolls down an empty city street at golden hour, then ollies over a crack in the asphalt. Realistic, warm colors.
Audio: wheels on asphalt, the pop of the board, distant traffic.

Realistic, warm color grade.
[00:00-00:02] Close-up of an alarm clock turning to 6:00; a hand slams it off.
[00:02-00:05] Medium shot: a young man sits up in bed and rubs his eyes in grey morning light.
[00:05-00:08] Wide shot: he pulls the curtains open and sunlight floods the small room.
Audio: alarm beep, a sleepy groan, curtains sliding, birds outside.

Dialogue and sound

• Veo 3.1 always makes sound with the video, so describe it or the model will invent it. Put sound in its own sentences: sound effects, background sound and dialogue.
• Name who speaks and keep the line short enough to say in the clip. Google's Flow tips warned that speech is less likely to come out when the line does not fit in an 8-second clip.
• Quotation marks or a colon? Google's own pages disagree. The newest Google Cloud guide (October 2026) says to skip quotation marks and write the line after a colon, as in: A woman says: My name is Clara. It warns that quotes can make the model write the words on screen. Older Google guides use quotes. Test both on one clip before a batch.
• To keep speech or music out, say so in a short line such as No dialogue or No music. Google's Gemini Omni Flash documentation suggests short lines like these, because it has no separate negative prompt.
• If Flow cannot produce good audio for a Veo clip, it does not create the video and refunds the credits.

One moment per clip, or timed shots

Google's best-practice guide says to keep each short clip to one focused moment. Do not chain several actions in one prompt; make separate clips and join them later.

When you do want several shots inside one 8-second clip, Google's Veo 3.1 guide suggests timed segments such as [00:00-00:02] followed by the shot, with sound notes per segment (see the last example above). Gemini Omni Flash uses a similar form, [0-3s] and so on, and also understands plain timing such as after 3 seconds.

Keeping a character the same across clips

• Use Characters in Flow. A character bundles one or two images, a name and a voice; type @ and its name in a prompt to use it.
• Use Ingredients for reference images of a person, object or place. Show subjects on a plain background and keep your text consistent with the images.
• Write one fixed character description (age, build, hair, face, clothing, voice) and paste it unchanged into every prompt. Change only the action and the setting.
• Use Frames to start a clip from a still image, for example the last frame of the previous clip, or to animate a photo.
• If Gemini writes several prompts for you, tell it to repeat every essential detail in each one.

Animating an image

When you start from an image, the image already shows what things look like. Google's advice is to prompt only for motion: how the camera moves, what the subject does and what moves in the background. Refer to the subject simply as the subject, he or she, and do not describe the picture again.

Slow push-in. The subject turns her head toward the window and smiles; the curtains move gently in the breeze.
Audio: soft wind, distant street noise.

What to avoid

• Several actions in one clip. Keep each clip to one moment, unless you use timed segments.
• Long, rambling description. Google recommends clear, direct sentences.
• Text that contradicts your references. The prompt should agree with the frames and ingredients you added.
• Negatives the model misreads. Describe what you want to see; when you must exclude something, use a short plain line such as No dialogue.
• Expecting JSON to work better. No Google guide documents a JSON prompt format. It works only as organized text, and plain sentences with the parts above do the same job.
• Prompts in other languages when English is an option. Flow Help recommends English prompts for the best results.

Running many prompts in a row

Flow takes one prompt at a time. For a video with dozens of scenes, ViralDNA Auto Flow, a free Chrome extension, sends your list of prompts to Flow one after another, one per line, and downloads every image and video in order, so scene 12 is file 12. It uses whatever model you pick in Flow. Its Img→Video tab turns a set of images into clips: Start gives one video per image, and Chain links the images into connected clips. Each clip still uses your Flow credits.

ViralDNA is not affiliated with Google. Google, Google Flow, Veo and Gemini are trademarks of Google LLC.