How to write cinematic Veo 3.1 prompts in batch

Most of the Veo prompts going around come back mediocre, because they skip four variables the model reads as directing instructions. The gap between an amateur prompt and a cinematic one is not talent — it is a reproducible framework. This guide is that framework: six explicit variables, eight worked examples, and how to apply it in batch with Flow AutoKuak so you are not typing the same scheme forty times by hand.

Why generic prompts fail

Take a concrete case. You ask Veo for “a woman walking on the beach at sunset”. Veo gives you something serviceable. You ask for “a woman walking on the beach at sunset, handheld camera”. Veo gives you something different, with more energy in the movement. You ask for “medium-long shot of a woman walking on the beach at sunset, handheld with soft steadicam, warm side-lit golden hour”. Veo gives you something that looks like an actual short film.

All three prompts describe the same subject. What changes is how many variables of film language you switched on. Veo 3.1 was trained on industrial quantities of annotated film and advertising — it knows what a dolly-in is, what anamorphic glass looks like, what backlight does. It just needs you to say so precisely.

The SCAPI framework: six variables that control the shot

The six variables any cinematic prompt should declare explicitly:

Variable	What it controls	Key vocabulary
Subject	What is on screen.	Subject + physical attributes + wardrobe + emotional attitude.
Camera	How the camera moves.	Static, dolly in/out, tracking, crane, handheld, drone, gimbal.
Action	What the subject does.	Concrete verb + direction + tempo (slow, abrupt, slow motion).
Plane	Shot type and framing.	Close-up, medium, medium-long, wide, high angle, low angle.
Illumination	Light and visual mood.	Golden hour, blue hour, cold neon, dramatic chiaroscuro, naturalistic.
Style (optional)	Visual or aesthetic reference.	Cinematic 35mm, anamorphic, Wes Anderson, A24, documentary realism.

Plus two modifiers that can travel on their own: duration (typically 4 or 8 seconds on Veo 3.1) and aspect (16:9, 9:16, 1:1). Those are set in Flow rather than in the prompt — but writing them into the prompt helps keep batches consistent with each other.

Fillable template

The working template for producing coherent prompts:

[SHOT TYPE] of [SUBJECT with concrete attributes], [ACTION with verb + tempo], camera [MOVEMENT TYPE], lighting [LIGHT DESCRIPTION], style [VISUAL REFERENCE]. Duration: 8 seconds. 16:9.

Filled in:

Medium-long shot of a 30-year-old woman in a flowing white dress, walking slowly towards camera with a melancholy expression, camera slow smooth dolly-out, warm side-lit golden hour with lens flares, cinematic 35mm style with subtle grain. Duration: 8 seconds. 16:9.

Compare that with the generic version (“a woman on the beach at sunset”). Five more variables, five specific orders to the model. The visual difference is enormous.

Eight worked prompts, for eight different intentions

1. Exterior establishing shot — cityscape

Aerial wide shot of a modern city skyline at blue dawn, low haze between skyscrapers, drone camera on a slow approach from 200 metres, cold blue-hour lighting with office lights still on, modern cinematic style like an HBO title sequence. Duration: 8s. 16:9.

2. Intimate shot — emotional close-up

Close-up on the face of a 50-year-old man, greying beard, looking off camera with glassy eyes about to well up, locked-off camera with a slight natural handheld tremor, soft Rembrandt side lighting with the background in deep shadow, anamorphic cinematic style with creamy defocus. Duration: 4s. 16:9.

3. Dynamic action — exterior chase

Medium shot of a runner in a red hoodie fleeing down a cobbled European street, handheld camera tracking laterally on a steadicam, flat overcast daylight with light mist, realistic cinematic style like Children of Men with controlled motion blur. Duration: 8s. 16:9.

4. Commercial product — cold drink

Macro shot of a soft-drink can covered in condensation, droplets rolling slowly, camera orbiting 90 degrees at medium speed, dramatic backlight with a blue side bounce, premium advertising style like an airline spot, black bokeh background. Duration: 4s. 1:1.

5. Documentary nature — wildlife

Medium shot of a red fox crouched in an autumn forest among crackling leaves, head raised alert towards a sound, static telephoto camera with slight slow motion, golden light filtering through trees like a forest at dawn, nature documentary style like BBC Planet Earth. Duration: 8s. 16:9.

6. Atmospheric sci-fi — ship interior

Medium-long shot of an astronaut in a clear helmet holding a glowing tablet in an empty spacecraft corridor, slow dolly-in towards the reflected face, blue-cyan neon lighting with screen flares, sci-fi cinematic style like Annihilation with subtle industrial grain. Duration: 8s. 16:9.

7. Quiet everyday — domestic scene

Medium shot of a 70-year-old woman drinking coffee by the window of an old kitchen in morning daylight, locked-off camera with centred symmetrical composition, warm side light coming through the window with visible floating dust, intimate cinematic style like Lost in Translation with a warm palette. Duration: 8s. 16:9.

8. Extreme action — slow-motion sport

Side shot of a skater doing a flip on a vertical ramp against a dusk sky, low-angle camera tilting up to follow the trajectory, dramatic backlight from the low sun with bluish anamorphic flares, extreme sports cinematic style like Red Bull, 240fps slow motion. Duration: 4s. 16:9.

---

Running the framework in batch with Flow AutoKuak

Writing 20 prompts in one sitting takes 30-45 minutes once the SCAPI template is clear in your head. Producing 20 Veo clips by hand in Flow takes 2-3 hours. With Flow AutoKuak in batch mode, those 20 clips come out in 35-50 minutes without you touching the interface. The flow:

1. Make a Google Doc with your 20 prompts. Separate each one with a clear delimiter (--- or ###END###).
2. Open labs.google/flow with your project active. Set the model (Veo 3.1 Quality or Fast), duration (4 or 8s) and aspect.
3. Open Flow AutoKuak from the side panel. Pick the source tab (Google Doc) and the target tab (Flow). Set the delimiters.
4. Run a pilot batch with the first 2-3 prompts (range 1-3 in the extension) to check the capture is clean before launching the full run.
5. Set a long timer (Veo Fast: 60-90s; Veo Quality: 120-180s) to give the render time between prompts.
6. Press Start. Clips download to your Downloads folder with sequential renaming (clip-001.mp4, clip-002.mp4, and so on).

Common mistakes that ruin the output

• Contradictions inside the prompt. “Wide shot with a close-up of the face” is a contradiction. Veo attempts both and returns something muddy. Pick one.
• A list of adjectives with no grammar. “Woman beach sunset handheld golden hour cinematic”. The model handles built sentences far better than keyword strings.
• Several subjects with no hierarchy. If the scene has three people, say who the protagonist is. “Three people talking” gives you three generic people; “close-up on the woman in the middle, two men out of focus either side” has a clear hierarchy.
• Styles that are too abstract. “Dreamlike style” says almost nothing. “Dreamlike style like Mulholland Drive with a saturated palette and gaussian blur” says a lot. Anchor the abstraction to references.
• Ignoring the camera. With no camera instruction, Veo defaults to static. For dynamic scenes, name the movement.

---

Iterating: when to accept a clip and when to rewrite

A realistic discard rate on Veo 3.1 sits around 30-50 % on new batches. Rules for deciding:

• Accept: the subject and action are right, the shot type matches, the light is coherent with what you asked for. Minor imperfections can be fixed in the edit.
• Rewrite the prompt and relaunch: the model misread the subject (a woman instead of a man, an age way off), the camera move is wrong (you wanted a dolly and got static), or the composition breaks the narrative.
• Switch to Veo Quality: the prompt is right but the output is noisy or loses detail. Going from Fast to Quality costs more credits, but the difference shows in complex scenes (crowds, water, fire).

If you want the exact credit economics per clip and when Quality beats Fast, read what a Veo 3 video actually costs in 2026. To avoid model-specific mistakes, the base Flow AutoKuak guide covers the whole flow. And to keep characters consistent between clips, how to keep character consistency in AI video is required reading.

---

Frequently asked questions

How many words should a Veo prompt have?

100-150 words (3-6 sentences) is the sweet spot. Under 50 loses control. Over 200 confuses the model and it starts ignoring instructions near the end. Google DeepMind documents this range in its official prompt guide.

Can I reference specific films or directors?

Yes, within limits. Style references (“like Wes Anderson”, “Children of Men style”) work. Asking it to reproduce specific scenes does not, and may breach the terms of use. Keep references as aesthetic guidance, not as copying.

Does SCAPI work on Veo 3.1 Fast too?

Yes. The prompt structure does not change between Fast and Quality — only the fidelity of the render. Veo 3.1 Fast costs 20 credits and is ideal for iterating on prompts before producing the final in Quality (50-100 credits).

How do I choose clip duration — 4 or 8 seconds?

4 seconds for one-beat shots (reactions, details, transitions). 8 seconds for scenes with development (action, exploration, narrative). For one-minute pieces, a 60/40 mix of 8s to 4s usually gives good rhythm.

Can I combine batch prompts with reference images?

Yes. Flow AutoKuak supports the image-to-video flow: you load an image as the visual reference and the text drives the animation. Useful when you already have a storyboard generated in Grok (Sequence mode) and want to turn it into Veo clips.

What is the prompt cap per batch?

On Free, 30 prompts per 24 h per tool (Flow AutoKuak counts separately from Grok and ChatGPT). On Pro, no cap per tool. The Veo model cap depends on your Google One credits — each clip costs 20-100 credits depending on the model.

---

Start with a batch of ten

The recommended first exercise: write 10 prompts applying SCAPI to your vertical. Run all 10 through Flow AutoKuak on Veo 3.1 Fast so iteration stays cheap. Look at which shots came out well and which prompts need rewriting. Once the 10 are tuned, relaunch only the problem ones on Veo Quality. That method — draft in Fast, final in Quality — is how you cut the discard rate and keep your credits. For how this fits the rest of a creator pipeline, read how to automate Grok (storyboards before Veo) and how to bulk-prompt ChatGPT (pulling shots out of a script).