AI Cinematic Video Prompts: Five Shot Recipes You Can Direct in Eight Seconds

Most short AI clips do not fail because the idea was too small. They fail because the prompt never told the camera what to do. Ask for "an epic cinematic shot of a musician" and you get a default wide frame with a slow drift, because every word in that sentence — epic, cinematic, amazing — names a feeling rather than a decision the model can execute.

Veo 3 raised the stakes for this kind of work. The model generates picture and sound together, producing dialogue, sound effects, and ambient noise natively rather than layering them on afterwards (Google DeepMind's Veo page describes it as video meeting audio). A single prompt can now return a line of dialogue with matching lip movement, footsteps on wet stone, and the room tone underneath. That turns prompting into something closer to a shot brief than a search query, and it means AI cinematic video prompts deserve the same discipline you would give a real shooting script.

This guide is about directing, not describing. It covers the structure that keeps a prompt readable to the model, the physical limits you have to design around, and five creative directions — a product reveal, a dialogue beat, an observed moment, a transformation, and a vertical hook — that you can build today on a prompt-first tool such as the Veo 3 generator on KOOX AI.

Why Cinematic Adjectives Do Nothing

Google's own guidance for Veo is blunt about ordering: cinematography comes first, because camera work is "the most powerful tool for conveying tone and emotion" (Prompt Architects' breakdown of Google's five-part formula). The five slots run cinematography, subject, action, context, then style and ambiance. When a slot is left empty, the model fills it with a default, and the default is always blander than your idea.

Named camera moves carry dense meaning. A "slow dolly push-in" tells the model the pace, the lens feel, and the emotional direction of the shot in three words; "cool camera movement" tells it nothing. The same logic applies to light. "Cinematic lighting" is a mood board; "hard key from camera-left neon, soft fill from a shop window on the right, low-key contrast" is an instruction a reviewer can actually check.

Google DeepMind's prompt guide works the same principle from the other direction: start by defining the sort of video you want to make — realistic, animated, stop-motion — and use character dialogue to set the tone. Style first, specific detail second.

The Six-Slot Shot Brief

Everything below fits one reusable template. Fill every slot that matters and leave the rest out; a silent product shot needs no dialogue, and a shot guided by a reference image does not need a paragraph about appearance.

• Subject — one person or object with one identifying detail. "A woman in her 40s, dark hair, wearing a linen blazer."
• Action — one visible event with a direction. "Walks slowly toward camera, glancing down once."
• Scene — place, time of day, weather. "A narrow bookshop at dusk, rain on the window."
• Camera — one shot size, one lens, one movement. "Medium close-up, 35mm, slow dolly push-in."
• Light — source, direction, quality. "Warm window light from screen left, soft fill, no haze."
• Audio — dialogue in quotes, effects and ambience named separately. "SFX: rain on glass, a door chime. She says: 'We're closed.'"

Two habits keep the template honest. First, budget the modifiers: two or three per shot is the sweet spot, and stacking five contradictory moves is the fastest way to a muddy result. Second, state the movement you want even when you want none — if you do not write "static" or "locked off," the model is free to add drift of its own.

A worked example, assembled:

Medium close-up of a tired baker leaning on a stainless counter after closing, 50mm, camera locked at eye level. Warm fluorescent light above, cool blue street light through the window, flour dust on the apron. She exhales and says quietly: "Tomorrow, we start earlier." Audio: refrigerator hum, one tray settling, faint rain. No music.

Design for Eight Seconds

Veo 3.1 generates eight-second videos with natively generated audio, at 720p, 1080p, or 4K (Gemini API Veo documentation). Eight seconds is not a limitation to fight; it is the unit of work. One shot, one readable event, one camera move. A beat sheet of three moments — establish, change, land — fits comfortably, and a prompt that asks for a minute-long oner will simply be cut short.

Length comes from assembly rather than a single generation. Extend continues a clip from its final second, so a sequence can grow to a minute or more while keeping visual and audio continuity (Google's Veo 3.1 announcement). First-and-last-frame generation bridges two images into a transition, and up to three reference images can anchor a character, product, or style when consistency matters. If you are starting from a still instead of a sentence, the same shot brief applies to image-to-video — describe what changes from the source frame rather than re-describing what is already visible, as in these photo-to-video workflows.

Five Directions You Can Direct Today

1. The product reveal with one material change. Pick a single product and one physical change: condensation appearing on a cold can, steam curling off a plate, liquid settling in a glass. Keep the camera static, or give it one short push-in that stops before the label fills the frame. Ask for one designed sound cue and room tone, and nothing else. The review test is simple: is the label still readable on the final frame?

2. The dialogue beat. One speaker, one short line, one locked medium or close shot. Put the spoken sentence in quotation marks so the model treats it as speech, name the delivery, and keep the line short enough to fit the clip. Check lip timing, pronunciation, and whether unwanted subtitles crept in before you judge the picture.

3. The observed documentary moment. Handheld follow or a slow track, one subject doing something ordinary and specific — a cyclist waiting at a red signal, a trader closing a stall. Keep the ambience layered but low: street hum, a bell, one contact sound. This is the direction that most rewards restraint, because natural motion and spatial logic are what make a clip read as real.

4. The surreal transformation. One object changes state inside a locked frame: a paper crane unfolding into a bird, sand rearranging into a face. Locked framing is essential, because the whole clip is the change itself. Give the moment material sounds instead of a score, and make sure the transformation is visible rather than implied.

5. The vertical hook. Immediate action in a vertical frame, one crisp effect in the first beat, the subject held inside the safe area. Portrait and landscape are both available, so decide the framing before you write, not after. Say "static" or request a single short push; a hook that drifts is a hook that loses its subject.

Each of these stays inside one eight-second event, which is exactly why they work.

Direct the Audio Like a Second Track

Native audio is the feature most people waste. Veo 3 generates dialogue, sound effects, and ambient noise from the prompt itself, and Veo 3.1 extended that audio support across the tools built on top of it. Three conventions keep the results clean:

• Dialogue goes in quotation marks and is attributed to a named speaker.
• Effects are named as effects — "SFX: a single aluminum slide, one drop hitting stone."
• Ambience gets its own sentence: "room tone, distant traffic, no music."

Silence is a choice worth writing down. "No dialogue, no music" is a legitimate instruction, and it is the honest way to keep a product or nature shot clean.

Build a Library, Review Every Take

Treat iteration as change-one-variable. If the composition is wrong, rewrite the camera slot; if the mood is off, rewrite the light slot; if the motion is wrong, keep everything else and swap the single move. Comparing two generations that differ in one field tells you which field caused the difference — and where you changed three things at once, you learn nothing.

With that comparison habit you can keep a small shot library keyed by intention — reveal, tension, transition, proof — each entry holding a fixed shot specification with variable content. Reuse the specification and swap the subject to keep a campaign visually coherent.

Two rules keep the practice clean. Generated clips carry a SynthID watermark (Veo 3 model card), so do not strip or misrepresent provenance, and disclose AI generation where your audience, a client, or a platform expects it. And do not prompt for real, identifiable people, protected characters, or branded products you have no right to use. A prompt brief is a plan, not a permission slip.

Start From One Line

The gap between a clip that looks generated and one that looks directed is usually two or three specific words: the name of a camera move, the direction of a light, the texture of a sound. Write the subject, give the camera one job, name your light, specify the sound, and design for eight seconds. Then generate, change one thing, and generate again.

Open the Veo 3 video generator, write your first shot brief, and let the camera do the acting.