This Veo 3 prompting guide turns Google's own Veo advice into a working method: the order to write things in, the camera words to use, how to cue dialogue and sound, and how to prompt from a start image. It applies to Veo 3.1 in Google Flow and the Gemini API. The same order of shot, subject, action, setting, style and sound is also a sensible way to write for Toybox AI's Video Creator. Video Creator is powered by Google's Gemini Omni Flash, not Veo, so the same prompt gives a different clip there. A short section at the end covers writing a separate soundtrack. Ready-made examples to copy are in the Veo 3 prompt library.
What Google's Veo 3 prompting guide recommends
Google's guidance describes one shot at a time, in a fixed order. An AI video prompt is the text description a video model turns into a clip, and Google's Veo 3.1 guide gives this formula for it: cinematography, subject, action, context, then style and ambiance. The DeepMind guide for Veo 3 lists the same ideas as seven parts: shot framing and motion, style, lighting, character, location, action and dialogue.
| Part | What to write | Example wording |
|---|---|---|
| Cinematography | Shot size and camera move | "Slow crane shot rising over..." |
| Subject | Who or what, with two or three specific details | "a teenage skateboarder in a faded orange hoodie" |
| Action | One thing that happens | "lands a kickflip and rolls away" |
| Context | The place and what's around it | "in an empty concrete skate park at dusk" |
| Style and ambiance | Look, light and mood | "warm backlight, 16mm film grain, nostalgic" |
| Sound | Speech, effects, background | "SFX: wheels on concrete. Ambient noise: distant traffic." |
Google's API documentation lists subject, action and style as the elements every prompt should include and marks the rest optional. Google's Veo 3.1 guide calls the camera work its strongest lever for tone and emotion, so write it first.
Writing camera direction for Veo 3
Camera direction goes first in the prompt, in the same words a film crew would use. Google's Veo 3.1 guide uses these terms in its examples:
| Term | What it looks like on screen | Good for |
|---|---|---|
| Dolly shot | The camera moves toward or away from the subject | Building attention on one detail |
| Tracking shot | The camera travels alongside a moving subject | People walking, cars, animals running |
| Crane shot | The camera rises or falls on a vertical path | Opening or closing reveals |
| Aerial view | Seen from high above | Landscapes and establishing shots |
| Slow pan | The camera turns sideways from a fixed spot | Rooms, skylines, crowds |
| POV shot | The view from a character's eyes | First-person moments |
| Arc shot | The camera circles around the subject | Product spins, dramatic reveals |
| Close-up, extreme close-up | The subject fills the frame | Faces, hands, textures |
| Low angle | The camera looks up at the subject | Making someone look tall or powerful |
| Two-shot | Two people in one frame | Conversations |
Lens words change the look as well. The same guide works with soft focus, deep focus, a macro lens, a wide-angle lens and shallow depth of field. Pick one camera move and one lens effect per clip; three moves in 8 seconds leave each one under 3 seconds.
How much fits in an 8-second Veo 3 clip
Each Veo 3.1 generation lasts 4, 6 or 8 seconds, and Google's API documentation requires the full 8 for 1080p, 4K, reference images and extensions. That length holds about one action, one camera move and one or two short lines of speech.
- One action. "Pours the tea" fits. "Pours the tea, sits down, opens a letter and starts crying" is a sequence of four clips.
- One or two lines. Read the dialogue aloud with a timer. If it takes more than about 5 seconds, cut it, because the character also needs a moment to react.
- Several shots with timestamps. Google's Veo 3.1 guide shows a timestamp format for a sequence inside one clip, with a bracketed time range in front of each shot, such as "[00:00-00:03]". Give each shot at least 2 seconds.
- Longer scenes through extension. Flow's Extend feature and the Gemini API continue a clip from its final second. Through the API, each extension adds 7 seconds, and a clip can be extended as many as 20 times.
Writing dialogue, sound effects and ambient noise
Veo 3.1 generates sound along with the picture, and Google's API guide says audio is always on. Sound goes in its own lines after the picture description, in three kinds:
- Dialogue. Google's API guide says to put specific speech in quotes. Name the speaker by appearance first: 'The older man in the gray cardigan says, "Leave the porch light on."' The DeepMind guide also shows a script style, with a name, a colon and the quoted line.
- Sound effects. Start the line with "SFX:" and name the sounds you'd hear, such as "SFX: a screen door creaks, then slams."
- Ambient noise. Start with "Ambient noise:" and describe the room or place: "Ambient noise: crickets and a far-off train."
Music has no separate label in Google's Veo 3.1 guide; its timestamp example puts an orchestral score in an SFX line. Write a score the same way: "SFX: a soft piano melody begins." Google DeepMind's Veo page lists natural-sounding speech in short lines as something it is still refining, so play back each line and regenerate a clip whose speech comes out garbled. For which video tools actually make clips with sound, see AI video with sound.
Negative prompts and things to avoid
Google's Veo 3.1 guide advises describing what the scene should contain instead of listing bans. Its example turns a bare "no man-made structures" into a landscape described as having no buildings or roads. So instead of "no people", write "an empty beach at dawn", and instead of "no text", write "a clean frame".
Video Creator has an optional "Things to avoid" field under More options. Even with it, say what fills the frame in the main description, the way Google's guide does.
Prompting from a start image
A start image sets the first frame, so the prompt only needs to describe what moves and how the camera behaves.
- Pick an image in the clip's shape. Use a vertical image for a 9:16 clip and a wide one for 16:9, so the framing doesn't shift.
- Describe the motion only. "The steam rises and the camera pushes in slowly" is enough for a photo of a cup of coffee.
- Say what stays put. "The mug and the table stay still" helps keep a product or a label steady, though no tool guarantees it.
- Add an end frame for a set landing. Flow's Frames mode and Toybox's Video Creator both accept a start and an end image, and the clip moves from one to the other.
- Use references for recurring characters. Flow's Ingredients and the API's reference images (up to 3) are Google's tools for holding a character or product steady from shot to shot. Video Creator's "Characters & objects" takes up to 3 photos of people, pets or products for the same job.
A Veo 3 prompt builder you can reuse
A prompt builder is a checklist you run through in order. Take a vague idea, "a video of a lighthouse at night", and build it one question at a time:
- Shot and move: a slow aerial orbit around the lighthouse.
- Subject details: a white lighthouse with a red band, on a rocky point.
- Action: the beam sweeps across the water as waves break on the rocks.
- Context: a stormy night, whitecaps, a fishing boat's lights far out.
- Style and light: cold blue moonlight, the warm beam, cinematic, deep focus.
- Sound: effects, background, and music if any.
The finished AI video prompt text reads like this:
Slow aerial orbit around a white lighthouse with a red band, standing on a rocky point on a stormy night. The lamp's beam sweeps across the dark water as waves break white against the rocks below, and a fishing boat's lights bob far out at sea. Cold blue moonlight against the warm beam, cinematic, deep focus. SFX: waves crashing, wind gusting. Ambient noise: a distant foghorn.Google's DeepMind guide says more detail gives you more control. The builder turns that into six answers, in order, in full sentences.
Prompting the soundtrack with Music Generator
When a clip needs a full song rather than a sound cue, write it as a separate job. Toybox's Music Generator writes an original song from two fields: "Song title" (up to 80 characters) and "Describe your song" (up to 1,000). A soundtrack description works best when it names four things:
- Genre and era: "lo-fi hip hop" or "1970s soul".
- Mood and tempo: "calm and slow" or "upbeat and bouncy".
- Instruments: "warm electric piano, soft brushed drums, upright bass".
- Where it plays: "background music under a cafe video, no big drops".
Choose "Instrumental" under "Song type" when the video has dialogue, so vocals don't compete with speech. Choose "With vocals" for a jingle, and paste your own words into "Lyrics" with tags such as [Verse 1] and [Chorus], or leave it empty and the lyrics are written for you. "Sounds to leave out" takes a short list such as "electric guitar, heavy drums". Each run makes two versions as MP3 files with cover art, currently costs 95 credits and needs a Pro plan. There's no length setting, so trim the song to the clip's length in your video editor.
Troubleshooting Veo 3 prompts
| Problem | Likely cause | Fix |
|---|---|---|
| The camera move never happens | The move is buried in the last sentence | Put the shot and move in the first sentence |
| The wrong character speaks | The line has no named speaker | Name the speaker by appearance before the quote |
| Speech sounds rushed or cut off | The line is too long for the clip | Cut it to one sentence you can say in about 5 seconds |
| The subject changes shape mid-clip | Too many actions for 8 seconds | Keep one action, or split the shots with timestamps |
| Words appear on screen | The prompt mentions signs, captions or quoted text you didn't mean as speech | Describe a clean frame, and quote only spoken lines |
| The start image changes too much | The prompt re-describes the whole scene | Describe the motion, and name anything that must stay put |
| It looks animated when you wanted real footage | The style isn't named | Open with "Realistic documentary footage" |
Run the same prompt in Toybox Video Creator
The lighthouse prompt's "Try this prompt" link opens Toybox's Video Creator with the text filled in. The six builder answers carry over as they are, but since Video Creator doesn't use Veo, expect a different clip than Veo 3.1 makes from the same prompt.
- Write the six builder answers as sentences in "Describe your video"; the box holds 2,000 characters.
- For image-to-video, attach the first frame with "Start from a photo", plus an end frame if the clip should finish on a set picture.
- Set "Format" to match the shape of any start image and "Length" to the seconds your action needs (4 to 10). Under "Quality", "Draft" (360p) is for testing the wording, and 720p or 1080p is for the clip you keep.
- Press "Create video". Expect a few minutes of waiting, and if the clip fails, the credits come back on their own.
- Check the clip against the builder: did the camera move happen, did one action play out, did the style hold.
- If one detail misses, describe the change in "Want to change something?" under the clip, or change one builder answer and run it again.
On a Pro plan or higher, a clip currently costs from 350 credits at 720p or from 450 at 1080p. Clips come with sound, so write the voices, effects and background you want into the prompt, with the same quoted lines and "SFX:" and "Ambient noise:" labels as above. Spoken lines may not come out word for word, so listen before you post. For a full song under the clip, drop in a Music Generator track while editing. For Google's own ways to reach Veo 3.1 and their free allowances, see how to use Veo 3.1 in Flow and Gemini.