VidoGridHow Veo 3 reads a prompt
Google splits a video prompt into subject, action, scene, camera, visual style and, from Veo 3 on, audio. Each part you leave out is a choice the model makes for you. Specific nouns help: 'a retired lighthouse keeper in a salt-stained wool sweater' gives Veo far more to work with than 'an old man'.
Write short, direct sentences rather than chatty descriptions. A dependable order is shot and camera move, then subject and action, then place and time, then style and light, then separate sentences for sound. Keep this Veo 3 prompt guide order as a template.
Slow arc shot around a matte black wireless speaker on a wet slate tile in a dark studio. A soft key light and a thin rim light trace its edges while water beads on the stone. Premium commercial look, cool grey palette. Audio: a low bass pulse that swells as the turn completes.
Camera, lens and pacing terms
Veo responds to ordinary film vocabulary. A low angle makes a subject look powerful, a high or top-down angle makes it small, and a Dutch angle adds unease. Framing runs from extreme close-up to wide establishing shot, and movement words include static, pan, tilt, dolly, truck, crane, aerial, handheld, whip pan and arc.
Lens terms such as telephoto, shallow depth of field, rack focus and lens flare change the look, while slow motion and time-lapse change how time feels. Google notes that some advanced angles and lenses are not officially supported, so test unusual terms on one sample first. Clips are short, so describe changes that fit in seconds, like a bud opening rather than a season passing.
Aerial drone shot gliding low over a braided glacial river at blue hour, slowly tilting up to reveal a jagged ridge with pink light on its snowfields. Deep depth of field, muted teal and rose palette, thin mist drifting over the gravel bars. Audio: steady wind and rushing water far below.
Sound effects, ambience and dialogue
Sound is the part of a Veo 3 prompt guide that silent video models never needed. Veo 3 and 3.1 generate audio together with the picture: sound effects tied to events on screen, ambient noise, and dialogue or voiceover. Say clearly that you want sound and give it its own sentences; DeepMind's examples often use an 'Audio:' label for this.
For speech, name the speaker, describe the voice (tone, pace, accent), then add a colon and the line. Google's best-practice page advises against quotation marks around dialogue, because quotes can make Veo draw the words as on-screen text. Keep lines short enough to fit the clip.
Medium shot in a cramped bicycle repair shop, late sun slanting through a dusty window. Maya, a mechanic in her forties with cropped grey hair and oil-stained overalls, spins a wheel and listens to it tick. In a dry, unhurried voice, Maya says: Your bearings are worn out, not your luck. Audio: the ticking wheel, a quiet radio, a passing tram.
Start frames, end frames, references and extensions
In image-to-video, your image becomes the first frame and already sets the subject, setting and style, so Google recommends prompting for motion only: camera, subject and environment movement. Call people 'the woman' or 'they' instead of re-describing them. Camera movement over a still scene is the most dependable option, and a sharp, well-lit source image gives a cleaner result.
All three Veo 3.1 models accept a first and a last frame, with the prompt describing the move between them. Generate and Fast take up to three reference images of a single person, character or product; on the standard model that mode makes 8-second clips only, and Lite does not support it. You can also extend a Veo clip of 1 to 30 seconds by 7 seconds; the extension picks up from the final second, so describe only what happens next.
Slow dolly in toward the subject. She turns toward the window and smiles slightly as a breeze lifts loose strands of her hair. Steam curls from the cup and the curtain behind her sways. Audio: quiet room tone, a kettle clicking off in the next room.
Settings that change the result
No Veo 3 prompt guide can fix a wrong setting, because some outcomes come from request parameters rather than words. Google documents these options for Veo 3.1 on the Agent Platform:
- Length: 4, 6 or 8 seconds, with 8 as the default.
- Aspect ratio: 16:9 (the default) or 9:16.
- Resolution: 720p by default or 1080p; the standard model's spec sheet also lists 4K output.
- Samples: 1 to 4 videos per request.
- Seed: the same seed and settings return the same video, so keep it fixed while you change one word at a time.
- Negative prompt: list unwanted things as plain nouns, such as 'captions, watermark, crowd', instead of writing 'no captions'.
- Person generation: allow_adult (the default) or disallow.
- Prompt language: English is the listed prompt language.
Veo 3 prompt guide checklist: mistakes and fixes
Run through this Veo 3 prompt guide checklist before spending a generation, and change one thing per attempt.
- Too many events in one clip. A few seconds hold one moment, so split 'unpacks the box, builds the shelf, hangs the photos' into three prompts.
- Dialogue showing up as captions. Remove the quotation marks and put a colon after the speaker.
- Re-describing the input image. In image-to-video, write only the motion.
- Characters drifting between shots. Name the character, paste the same full description of face, hair, build, clothing and voice into every prompt, edit only the action and setting, and reuse the seed.
- Using 'no' in the negative prompt. List the unwanted items themselves.
- Starting from a blank page. Both Google guides suggest asking Gemini to expand a rough idea into a detailed prompt.