The Veo 3 Prompt Structure: 6 Parts That Make It Obey

Veo 3 obeys structure, not enthusiasm. A prompt that clearly specifies subject, action, scene, camera movement, lighting, and style/audio — in roughly that order, written as natural sentences — will beat a pile of buzzwords every time. Below: each of the six parts, bad-vs-good examples you can copy, and what predictably goes wrong when a part is missing.

Why structure matters more than adjectives

Veo 3 (and its faster sibling, Veo 3 Fast) was trained on video paired with descriptive, caption-like text. That means it responds to prompts that describe a shot the way a director or cinematographer would, not to SEO-style keyword lists. "Cinematic, 4K, masterpiece, trending" tells the model almost nothing. "A slow dolly-in on a fisherman's weathered hands as he ties a knot at dawn" tells it everything: who, what, how the camera moves, and when.

The six-part structure isn't a magic template — it's a checklist. You don't need rigid labels or a specific order; you need each element present, because whatever you leave out, the model decides for you. And the model's default decisions are exactly what make AI video look generic.

The 6 parts, one by one

1. Subject — who or what we're looking at

Be concrete and visual. "A woman" gives the model a coin flip on age, wardrobe, ethnicity, and mood. "A woman in her 60s with silver hair, wearing a mustard raincoat" locks the character. If your subject is a product or object, describe material and scale: "a matte-black ceramic coffee mug" beats "a mug."

2. Action — one clear beat

Veo 3 clips run about 8 seconds as of this writing. That's one action beat, not a scene. "She picks up the mug, blows on the coffee, and smiles" works. "She wakes up, gets dressed, drives to work, and argues with her boss" does not — the model will smear four actions into visual mush or drop three of them.

3. Scene — where and when

Location, time of day, weather, and era. The scene anchors everything else: lighting becomes believable, ambient audio becomes predictable, and background details stop hallucinating. "A cramped Tokyo ramen bar at night, steam rising, rain visible through the window" does more work than any style keyword.

4. Camera movement — how the shot behaves

This is the part most beginners skip, and it's the part that most separates amateur output from professional-looking clips. Use real cinematography vocabulary: dolly in, slow push-in, tracking shot, handheld, crane shot, orbit, static wide, rack focus, low angle, POV. If you don't specify, Veo 3 tends to default to a slow, aimless drift that screams "AI video." It's the one default I'd pay Google to remove.

5. Lighting — the mood engine

Name the light source, quality, and direction: "soft golden-hour sunlight from camera left," "harsh fluorescent overheads," "single neon sign casting magenta on wet pavement." Lighting language is the cheapest way to buy production value; the model understands it remarkably well.

6. Style and audio — the finishing layer

Style: film stock references, genre look, color grade ("shot on 35mm, muted teal-and-orange grade," "1970s documentary look"). Audio: Veo 3 generates synchronized sound, so describe it. Put dialogue in quotes, name ambient sounds, and specify tone of voice. If you don't mention audio, you get whatever the model guesses — often awkward murmurs or silence.

Bad vs good: three worked examples

Example 1 — character clip

A cool cyberpunk girl walking in a city, cinematic, 4k, ultra detailed, epic lighting, masterpiece

This is a keyword pile. No action beat, no camera, no light source, no audio. You'll get a generic drifting shot of a generic character.

A young woman with a silver bob and a translucent rain poncho walks toward camera through a crowded neon market street in Osaka at night. Slow tracking shot moving backward at her pace, waist-up framing. Rain-slicked pavement reflects pink and cyan signage; her face is lit by a passing storefront. Ambient audio: rain, crowd chatter, distant synth music from a shop. Shot on anamorphic lenses, shallow depth of field.

Every one of the six parts is present, so the model has zero important decisions left to improvise. Swap the city, wardrobe, or color palette and the skeleton keeps working.

Example 2 — dialogue clip

A chef talks about his restaurant, make it emotional and cinematic

"Talks about" with no actual line means Veo 3 invents mumbled non-language or drifting lip movement — a classic failure.

Medium close-up of a gray-haired Italian chef in his 60s standing in a small trattoria kitchen, warm tungsten light from a single overhead lamp. He looks slightly off-camera and says warmly: "My father opened this place in 1962. Every plate still comes from his recipes." Static camera on a tripod, shallow focus, gentle kitchen ambience — a pot simmering, cutlery in the distance. Documentary interview style.

The quoted line gives the audio model something concrete to lip-sync, and "static camera, documentary interview style" kills the drifting-camera default. Change the quote and location to reuse this for any talking-head clip — there are more patterns like this in our product ad prompt pack.

Example 3 — action clip

An epic car chase at night with explosions and drifting and helicopters, super fast and intense

Four competing subjects and no single action beat. Expect physics soup.

A black 1970s muscle car drifts around a rain-soaked corner in an empty downtown street at night, tires smoking. Low-angle static camera close to the asphalt as the car slides past, headlights flaring the lens. Sodium-vapor streetlights, wet reflections. Audio: engine roar, tire screech, echo off buildings. Gritty 35mm action-film look.

One car, one corner, one camera position. The intensity comes from the low angle and lens flare, not from stuffing more chaos into eight seconds.

What breaks when each part is missing

This table is the whole article in one glance. If a generation looks wrong, find the symptom here and you'll usually find the missing part.

A reusable template

Keep this skeleton in a notes file and fill in the brackets:

[SHOT TYPE] of [SUBJECT with 2-3 concrete visual details] [doing ONE action] in [LOCATION, time of day, weather]. [CAMERA MOVEMENT]. [LIGHTING: source, quality, direction]. Audio: [dialogue in quotes and/or ambient sounds]. [STYLE: film stock, genre, grade].

It reads like a shot list because that's exactly what it is. Once this becomes muscle memory, you stop re-rolling generations and start directing them.

Common questions about ordering and length

Does the order of the six parts matter? Somewhat. Veo 3 weights earlier tokens more heavily, so lead with subject and action — the things you care most about — and finish with style. Prompts in the 40–120 word range perform best in my testing; below that the model improvises too much, and far above it instructions start to contradict each other.

One more habit worth building: change one variable at a time between generations. If you rewrite the whole prompt after every result, you never learn what actually moved the output. Treat it like a science experiment, not a slot machine. For the mistakes that cause most bad generations, see why your Veo 3 videos look bad, and when you're ready to copy proven prompts, start with the 15 cinematic Veo 3 prompts.

If you want new tested prompt structures like this in your inbox each week, join the free newsletter.

Keep reading