Google Flow Prompt Guide: How to Prompt Omni and Veo (2026)

A good Google Flow prompt reads like a shot description. It names the subject, the action, the environment, the lighting and the style, and then it adds the camera move and the sound you want. As of November 2026, Flow's video models are Omni Flash 1.2 and Veo 3.1 Lite, and they reward different habits: DeepMind says Veo needs precise instructions, while Omni needs less prescriptive prompts. The cheapest way to learn is to draft at 360p, where Omni Flash 1.2 costs 1 credit per second, and to fix a clip with one targeted edit instead of starting over. Everything below comes from Google's own prompt guides and Flow Help; every example prompt is ours.

Choose the model first, then write for it

Prompting starts with a model choice, because the two models in Flow accept different inputs. Google retires Veo 3.1 Fast, Veo 3.1 Quality and Omni Flash 1.1 in Flow on December 5, 2026, and until then they stay available at standard credit rates; this guide covers the two models that remain. Omni Flash 1.2 handles Text, Frames, Ingredients and video-to-video editing, at 3 to 10 seconds for everyone and up to 15 seconds on paid plans. Veo 3.1 Lite makes 4, 6 or 8 second clips at 720p, runs Ingredients at 8 seconds only, and cannot edit or extend video.

The prompting difference is the useful part. DeepMind's Omni guide says Veo works best with precise instructions, while Omni's reasoning and world knowledge let you write less prescriptive prompts. In practice, a Veo 3.1 Lite prompt should spell out each element, because a fixed 10-credit clip leaves no room for follow-up edits. An Omni prompt can start looser and be refined afterward. Pick Veo 3.1 Lite for a single, fully specified shot. Pick Omni when you expect to iterate. Our Veo vs Omni comparison covers costs and the retirement dates in depth.

The five elements Flow Help asks you to name

Flow Help gives one core rule: describe the scene in detail and be specific about the subject, action, environment, lighting and style.^[1] DeepMind's Veo guide expands that list to shot framing and motion, style, lighting, character descriptions, location, action and dialogue.^[2] Two habits from that guide pay off quickly. Start the prompt by naming the kind of video, such as realistic, animated or stop-motion. Then describe the location thoroughly; the guide's own contrast is a jazz club versus a smoky jazz club at night.

Here is how we would assemble those parts. This example is ours, not Google's: Realistic handheld footage. A barista in a green apron pours latte art into a white cup on a wooden counter, early morning, soft window light from the left, warm muted colors, shallow depth of field. The camera slowly pushes in on the cup. Quiet café ambience, a steam wand hissing. Each sentence answers one of the elements, so nothing is left for the model to guess.

For fast action, the Veo guide recommends mapping out an exact play-by-play. For everyday subjects, it suggests a clear beginning, middle and end. Both ideas fit Flow's short clips well. Our rule of thumb: three beats per clip. Omni clips top out at 10 seconds without a plan and 15 on paid plans, so a story with more beats than that belongs in two clips joined in Scenebuilder. Split it instead.

Camera language and on-screen text

Camera direction is the easiest place to be vague. DeepMind's Omni guide lists the vocabulary it responds to: one continuous shot or a oner, a static or locked-off frame, a push in, a punch in, a dolly zoom, a natural smartphone zoom, film-camera looks and webcam style.^[3] Use the term itself rather than describing the effect. Locked-off wide shot is clearer than the camera stays still and shows everything, and it leaves fewer words open to interpretation.

Style can change without changing motion. The same guide shows Omni restyling a reference video as anime, claymation or watercolour while keeping the original movement. That makes a useful workflow: generate the motion you like in a realistic style, then ask for the restyle as a separate step. Omni can also render text in sync with the picture, and the guide lets you specify the type, placement, animation and exposure. Spell out each of those four. A title that only says add a headline leaves the font, position and timing to chance.

For longer ideas, the Omni guide describes sharing a visual storyboard and asking the model to follow its beats in order, for example from the top-left panel. Flow also ships a Google-made Storyboard Studio Tool for writing a script and visualizing a storyboard. Both are worth using before you spend credits on finished clips. Plan first. Then generate.

Dialogue and sound need their own lines

Audio is generated with the video, so it belongs in the prompt. DeepMind says Veo generates sound effects, ambient noise and dialogue, and the Veo guide gives two ways to direct speech: name a topic for the characters to discuss, or give them exact lines. It also advises defining the sounds explicitly, either inside the prompt or in a separate section. Give sound its own final sentence, so it never gets lost among visual details.

Our own example of that structure: A woman at a bus stop checks her watch and says, I'm going to be late again. Sound: light rain on the shelter roof, a distant bus engine approaching. We suggest one short line per speaker rather than a speech, especially since DeepMind itself notes that natural spoken audio for short speech segments remains an area of active development. Expect some retries.

Two Flow settings change what happens to audio. If Veo produces poor audio, Flow does not return the video and refunds the credits, unless the project setting Return silent videos is on; then you get a silent clip and pay for it. On Omni Flash 1.2, voice references, a single speaker referenced with @Voice, work only in generations that use ingredients. You can also build a custom voice by choosing a preset and describing how to restyle it. Some voice options are limited to certain regions.

Frames, ingredients and @characters

Once you add images, the prompt's job changes from describing to supporting. Flow Help says the text should complement, not contradict, your frames or ingredients, and that subject or product references work best on a plain or segmented background. A prompt that describes a red jacket over a reference showing a blue one gives the model two instructions. Keep them aligned.

For first and last frames, DeepMind's Omni guide suggests three parts: define what the first frame shows, specify the exact camera movement or transition between them, and say what the final frame reveals. Orbits, zoom transitions and loops are the guide's examples. Frames to Video with Omni Flash 1.2 runs on the web only, not in the mobile app. Our image-to-video guide walks through the frame options step by step.

Characters keep a face, clothing and voice consistent across clips. A character needs one or two images plus a voice, and you call it in a prompt with @ and its name; your own avatar is @me. Avatars are not available in the EEA, the UK or Switzerland. DeepMind adds that references created with Nano Banana, Flow's image models, can keep characters, objects or environments consistent. One firm limit applies everywhere: Flow does not generate prominent people.

Iterate with edits, not fresh prompts

The biggest saving comes from how you revise. Omni Flash 1.2 at 360p costs 1 credit per second and at 720p 3 credits per second, for clips up to 10 seconds.^[4] By our arithmetic, one 10-second clip at 720p costs 30 credits, while the 50 free daily credits cover five 10-second 360p drafts. Drafting at low resolution until the prompt is right is simply cheaper. Resolution can wait.

Once a draft is close, change one thing. DeepMind's Omni guide recommends asking for a specific update, such as a new background or caption, without re-prompting the whole scene. Flow keeps the context of earlier edits for up to three conversational turns, and edits cost the same per second as generation. Extension is a separate step on paid plans only: it adds 3 to 10 seconds with up to three image references, since Veo Extend was retired on November 5, 2026. Other edit modes cannot be applied to an extended clip, so finish insert, remove and camera edits first.

Two helpers save typing. The Veo guide notes that Gemini can expand a short prompt with more detail, and the Flow agent can create several variations from one request. Credits are charged per generation, not per request, so a batch of variations costs a batch of clips. Unused credits never roll over; our credits guide explains the daily and monthly refresh.

The bottom line

Prompting Flow well is mostly discipline. Name the five elements Flow Help lists, use real camera terms, give sound its own sentence, and make the text agree with any frames or characters you attach. Match the habit to the model: precise, complete prompts for Veo 3.1 Lite, looser starts and targeted edits for Omni Flash 1.2. Draft at 360p, change one thing per turn, and save 720p for the version you keep.