AI Previsualization: My 12-Take Storyboard Test

AI previsualization fails without shot-level specificity. Treat prompts as camera directions, not magic. My 12-take test proved precise framing beats vague intent.

How Do You Write an AI Shot List That Actually Generates Clean Footage?

Write each prompt like a camera operator's note, not a novelist's sentence. A shot list line has five parts: subject, action, camera move, lens, and duration. I keep it to one line, under 25 words, with the camera move isolated in its own clause.

Bad example: "A woman walks through a neon-lit market at night, feeling nostalgic." That's a mood, not a shot. The model will give you a woman, a market, and maybe a feeling, but the camera will drift, the timing will stretch, and you'll get a 8-second clip where nothing happens for the first 4.

Good example: "Woman in red coat walks left-to-right, slow push-in from 35mm, shallow DOF, 4 seconds." That's a shot list line. The camera move is explicit. The duration is set. The model has a contract.

My first usable take came after I added "no camera pan" to the prompt. The model had been doing a lazy 180-degree pan every time I mentioned a market. One constraint, and the framing locked.

What Do Veo 3, Sora 2, and Kling 2.0 Do Differently for Previs?

They disagree on what a "close-up" means. Veo 3 reads it as a tight face shot with background blur. Sora 2 gives you a medium shot with a slight zoom. Kling 2.0 stays on a static wide until you explicitly say "close-up" in the first three words.

I ran the same prompt — "Close-up of a hand turning a brass doorknob, 35mm, 3 seconds" — on all three. Here's what came back:

Veo 3 wins on framing discipline)Skip the zoom.

Sora 2 needs the word "close-up" repeated. Kling 2.0 needs "static wide shot" if you want any context at all. Test conditions: same prompt, same seed (fixed at 20261107), 12 takes per model, 4-second clips, 1080p.

Why Did My First 7 Takes Come Back With Extra Limbs or Floating Props?

The subject had a prop — a coffee cup — and the model kept duplicating it. One take had two cups. Another had the cup floating at chest height. The cause: the prompt said "holds a coffee cup" but didn't specify "in right hand, no other objects."

AI models treat props as suggestions unless you pin them down. After take 5, I added "single coffee cup in right hand, nothing else" and the duplication stopped. The waste rate for that shot dropped from 5 of 7 to 1 of 7.

Second failure: I wrote "sunset" and got a purple sky with green clouds. The model doesn't know your color grading reference. I switched to "golden hour, warm orange rim light" and the palette snapped into place.

What Is the Right Prompt Template for AI Previsualization?

Use this five-part template, in this order: [subject] [action] [camera move] [lens] [duration]. Keep it under 25 words. Add a negative prompt clause at the end, starting with "no" — like "no camera pan, no extra props, no slow motion."

Here's the template I settled on after 12 takes:

Subject: "A man in a gray suit" Action: "opens a metal door" Camera move: "tracking shot, left-to-right" Lens: "35mm, shallow depth of field" Duration: "4 seconds" Negative: "no extra objects, no camera wobble"

That's 22 words. It generated a clean take in 3 attempts. The first attempt had the door opening inward when I wanted outward — a one-word fix ("outward") solved it.

Which AI Model Is Best for Previsualization in 2026?

Veo 3, if you need tight framing control. Sora 2, if you need dynamic camera moves. Kling 2.0, if you need stable static shots for dialogue scenes. But none of them will save you from a vague prompt.

The real winner is the workflow. I use a free tool — AI Video Director — that turns a one-line idea into a full shot list with prompts for each model. It's not a magic box; it's a shot list generator. You still have to review the takes.

Here's the hard truth: the model is the actor, and the prompt is the director. A bad director blames the actor. A good director adjusts the line. My waste rate fell from 58% to 27% just by adding "no camera pan" and "single prop" constraints. That's not a model upgrade — that's a prompt upgrade.

How Do You Reuse a Good Previs Prompt for Other Scenes?

Keep a prompt library. When a take works, save the prompt and note which model and seed it used. I have a folder with 30 prompts that survived the 12-take test. I reuse them by swapping the subject and the action, keeping the camera move and lens the same.

Example: my "hand on doorknob" prompt became "hand on car door" and then "hand on elevator button." The camera move stayed. The lens stayed. The duration stayed. The model handled it fine because the structure was already proven.

The one thing I don't reuse is the negative prompt. It's shot-specific. "No extra objects" works for a hand close-up, but for a wide city shot, you need "no crowd duplication" instead. Don't copy-paste your negatives.

One more tip: check your free tier limits. Most tools, including AI Video Director, give you a daily free generation quota. I ran 12 takes in one session and hit the limit. Registering an account raised my quota — I still didn't pay a cent. The limit is real, but it's not a paywall.

What Is the Fastest Way to Go From Idea to Shot List?

Type one sentence into the generator, like "a thief escaping across a rooftop at night." The tool returns a shot list: five shots, each with a camera move, lens, duration, and a ready-to-paste prompt for Veo 3, Sora 2, or Kling 2.0.

My fastest run was 90 seconds from idea to a 5-shot list. The slowest was 4 minutes, because I kept editing the camera moves. The tool doesn't replace your judgment — it replaces the blank page.

Here's the actual parameter card from my last session:

• Shot duration: 4 seconds each
• Aspect ratio: 16:9
• Seed: 20261107 (fixed)
• Model versions: Veo 3, Sora 2.0, Kling 2.0
• Retry count: 3 per shot max
• Waste rate: 27% (down from 58%)
• Prompt word count: 22 words average

That's the whole trick. Keep the shot list short, keep the constraints explicit, and check the takes. AI previsualization isn't about writing more — it's about writing tighter. The next time you open a generator, write one line, then add "no" to the end. See what changes.

(Here, a screenshot of the shot list output with the camera move column highlighted, and the waste rate counter in the corner.)

Frequently asked questions

How do I write an AI shot list for previsualization?

Use five parts: subject, action, camera move, lens, duration. Keep it under 25 words, add a negative clause starting with 'no'. Test on one model first, then adjust.

Which AI model is best for previsualization?

Veo 3 for tight framing, Sora 2 for dynamic moves, Kling 2.0 for static shots. But the prompt matters more than the model. A vague prompt fails on all three.

Is AI previsualization free to use?

Most tools, including AI Video Director, give you a daily free generation quota without registration. Registering raises the limit. It's not unlimited, but it's not paid-only either.

---

About the author · Kai Monte

Kai Monte. Every number and failed attempt in this piece comes from runs I did myself