How I Learned to Write AI Video Prompts (After Burning a Lot of Credits)

Two hundred burned credits later, I realized the models weren't broken. I was just writing like a customer texting a friend instead of a director on a film set.

When you type a lazy sentence, press generate, and get garbage, the instinct is to switch models. I cycled through three different tools before figuring out that the model was fine—my prompt was just making it guess everything.

No theory here. Just what actually works. If you want to follow along, open the Video Studio in another tab and try these as we go.

Stop making the model guess

A video model only knows what you explicitly tell it.

Here is what I used to type:

a woman walking in a garden

Six words. The model has to guess the garden, the time of day, the camera angle, her expression, and whether the camera is moving. That is twenty arbitrary decisions, and every guess is a chance to disappoint you.

Here is the same idea, written properly:

Medium tracking shot, camera moving slowly beside a woman in her
thirties walking through an overgrown summer garden. Warm golden hour
light coming from behind her, soft haze in the air. Shallow depth of
field, 50mm lens look, background blurred. Calm and thoughtful mood.
She brushes a hand across tall grass as she walks. 6 seconds, steady
natural pace.

Same subject, completely different output. I didn't get better at art; I just stopped leaving vacancies in the prompt.

The six-point checklist

I keep this in my notes and run down it for every render. If I skip one, that is usually the element that comes back wrong.

1. Subject and single action. Pick one moment. "She pours coffee, looks out the window, and the cat jumps on the table" is three separate prompts fighting over five seconds. Keep it to one clear move.

2. Framing and angle. Wide shot, medium, close-up, or extreme close-up. Eye level, low angle, high angle, or over-the-shoulder. Adding this single line stops your clips from looking random.

3. Camera movement. Static shot, slow push-in, pull-back, tracking, pan, handheld follow, or crane shot. If you leave this out, the model usually defaults to a lazy, slow zoom. If you want the camera to stand still, write "static shot, locked off camera."

4. Lighting. Light is 90 percent of what separates a cinematic shot from a cheap screen saver. Golden hour, blue hour, overcast, harsh midday sun, neon reflections, or a single lamp in a dark room.

5. Lens and style. 24mm wide for scale, 50mm for natural sightlines, 85mm for portraits with soft backgrounds. Add depth of field, film grain, or clean digital finish. If you use one of our studio presets (Cinematic, Anime, Realistic, Cyberpunk), keep the text prompt aligned so you aren't fighting the style.

6. Duration and speed. Stick to 5 to 10 seconds. Longer clips tend to drift—faces morph, hands melt. Specify the pace too: slow motion, real-time, or energetic.

Two to four sentences total. You don't need a massive wall of text.

Matching the model to the job

I treat models like different lenses now. Here is how I split them up in our studio:

Keep in mind that Veo likes structured, technical language, while Wan prefers plain visual description. Give a prompt one pass in the model's preferred style before judging the output.

Draft cheap, finish expensive

Don't iterate on top-tier models.

Run the initial idea through Wan 2.5 at 5 seconds. Render it five or six times, changing one variable per round—camera angle first, then lighting, then framing. It costs almost nothing per attempt. Once the composition actually matches what you envisioned, move that exact prompt over to Veo 3.1 or Seedance 2.0 for the final render.

Use the cheap model to debug the sentence, and only change one word at a time so you actually know what caused the shift.

Frame negatives as positives

Avoid negative prompt lists like "no blur, no distortion, no extra fingers." Models handle positive instructions much better than prohibitions.

Instead, write: "sharp focus throughout, clean stable hands, smooth continuous movement." Tell the model what to draw, not what to ignore.

Reference stacking limits

Style references work well ("Shot like a Wong Kar-wai interior, saturated reds, slow shutter smear"), but cap it at two references.

If you blend too many conflicting styles—like "Wes Anderson symmetry meets found footage horror"—the model averages them out into a muddy mess.

Story vs. Visuals

• Describe faces, not emotions: "She realizes he lied" is a script line a model can't render. Write "her smile fades, eyes drop, jaw tightens slightly."
• One shot per clip: Five seconds is a single shot. If you need three shots, generate three separate clips and cut them together.
• Audio cues: Seedance 2.0, Veo, and Wan 2.6 generate sound. Include half a sentence like "ambient wind and distant traffic, no music" to steer it.

How we built the Studio around this

We built OpenGradient's Video Studio specifically to fix the friction in this workflow.

Your prompts shouldn't be logged or tied to a permanent profile. When you generate clips here, the requests go through an anonymity layer—encrypted on your device, decoupled from identity via relays, and decrypted only inside attested enclaves. Share links automatically expire in 24 hours.

Instead of monthly subscriptions, it runs on standard pay-as-you-go credits (1,000 credits for $1), so you can burn through early test iterations on cheap models without wasting money.

The basic template

[Shot type] of [subject] [one clear action], [camera movement].
[Lighting] light, [mood/atmosphere].
[Lens] look, [depth of field], [film style].
[Duration], [pacing]. [Audio cues.]

Drop that into Wan 2.5, run a few low-cost passes to lock in the motion, then switch the picker over to Seedance 2.0 or Veo 3.1 for the final output.

You can try it out directly in the Video Studio.

---

*Model names belong to their respective creators (Seedance/ByteDance, Veo/Google, Kling/Kuaishou, Grok/xAI, Wan & Happy Horse/Alibaba open weights). OpenGradient is an independent platform.*