Sora 2 & Veo 3 Prompting Guide: How to Write AI Video Prompts That Actually Work

Why Sora 2 and Veo 3 need different prompts than image models

Image models like Midjourney and Nano Banana describe a single frozen moment. Video models like OpenAI Sora 2 and Google Veo 3 have to plan motion, physics, lip sync, and temporal consistency across 5–60 seconds. That changes how you write the prompt: you are no longer describing a picture — you are directing a shot.

The 7-block prompt formula that works for both Sora 2 and Veo 3

1. SHOT — frame size and angle (wide, medium, close-up, low-angle, top-down).
2. SUBJECT — who or what, described with concrete physical detail.
3. ACTION — one verb. Avoid chaining actions in a single clip.
4. CAMERA — lens, movement, speed. One move per clip.
5. LIGHTING — key direction, quality (soft/hard), color temperature.
6. ENVIRONMENT — location, time of day, weather, background depth.
7. STYLE — film stock, grade, era, reference (e.g. Kodak 500T, A24, anime cel).

Camera movement: what Sora 2 and Veo 3 can actually do

Both models render simple, named camera moves reliably. Stick to one move per clip and use the standard film vocabulary — the models were trained on it.

• Dolly in / out — physically moves toward or away from subject. Best for emotional emphasis.
• Track / parallax — moves sideways past subject. Great for reveals and landscapes.
• Orbit — circles the subject. Best at slow speed; fast orbits warp geometry.
• Crane / boom — vertical lift. Use sparingly; combine with locked horizon.
• Locked tripod — no movement. The most stable, highest-fidelity option. Underrated.

Avoid: "dolly in then crane up then orbit" in a single prompt — Sora 2 and Veo 3 will average the moves and you'll get drift and morphing geometry. One move per shot.

Temporal consistency: keeping the same face, outfit, and place across shots

The single biggest weakness of current AI video is identity drift between clips. Use these rules to lock it down:

1. Repeat the subject description verbatim in every clip prompt — same age, hair, eye color, wardrobe, ethnicity, accessories.
2. Anchor the lens — keep the same focal length (e.g. 35mm) across the sequence.
3. Anchor the lighting direction — if key is camera-left in shot 1, keep it camera-left in shot 2.
4. Anchor the color grade — name the same film stock or LUT in every prompt ("Kodak 2383", "teal-orange", "bleach bypass").
5. Use Sora 2's "remix" or Veo 3's "extend" instead of regenerating from scratch when you need continuity.

Sora 2 specifics

• Prefers single-paragraph natural language over bullet structure.
• Sora 2 reads audio cues — name dialogue, sound effects, and music mood inside the prompt.
• Best clip length: 8–12 seconds. Longer clips lose subject coherence.
• Strong at physics (water, cloth, hair) when you name the material explicitly.

Veo 3 specifics

• Prefers block-structured prompts (SHOT / SUBJECT / ACTION / CAMERA…).
• Veo 3 has native synchronized audio — always include an AUDIO line.
• Strongest at talking heads and lip sync — name the dialogue in quotes.
• Best at locked tripod and slow dolly; less reliable on fast handheld.

30+ ready-to-copy prompt templates

Below are battle-tested templates. Swap the subject and grade to fit your project.

Cinematic medium shot of a young woman standing on a rain-soaked Tokyo street at night, neon reflections on wet asphalt. Slow dolly-in, 35mm lens, shallow depth of field. Soft key light from the left, magenta and cyan rim light. Film grain, Kodak 500T look, 24fps.

SHOT: Macro top-down
SUBJECT: A glass bottle of cold brew coffee with condensation
ACTION: A single drop slides down the bottle
CAMERA: Locked tripod, 60mm macro, f/2.8
LIGHTING: Soft window light from the right, deep shadow on the left
ENVIRONMENT: Dark walnut table, scattered coffee beans
STYLE: Commercial, photoreal, muted earth tones
AUDIO: Subtle ambient room tone, soft drip

Wide aerial parallax shot tracking left to right over a misty pine forest at sunrise, distant mountains revealed behind the trees. Smooth drone motion, 24mm lens, golden warm grade, volumetric god rays through fog.

Medium close-up of a 40-year-old male presenter in a navy suit speaking directly to camera in a modern glass-walled office. Locked tripod, 50mm lens, eye-level. Soft key from camera-left, subtle fill, slight background bokeh. Natural lip sync, calm confident tone.

Low-angle tracking shot following a parkour runner leaping between rooftops at dusk. Handheld stabilized camera, 18mm wide lens, fast shutter, dynamic motion blur on edges. Warm orange sky, cool blue shadows, gritty cinematic grade.

Top-down 360° orbit around a steaming bowl of ramen on a dark slate plate. Slow rotation, 50mm, soft overhead window light, visible steam rising, chopsticks resting on the rim. Photoreal, restaurant commercial look. Seamless loop.

Common mistakes that break Sora 2 and Veo 3 outputs

• Too many subjects. More than 2 named characters → identity collapse.
• Chained camera moves. One move per clip, always.
• Vague style words. "Cinematic" alone means nothing. Name the film stock or director.
• Conflicting lighting. "Golden hour at midnight" — pick one.
• Over-long prompts. Past ~120 words, Sora 2 starts ignoring tail tokens.

FAQ