Every AI video starts the same way: a blinking cursor in an empty prompt box. You type "a beautiful sunset over the ocean," hit generate, and get… a sunset. Technically correct. Completely forgettable.
The difference between that clip and one that stops people mid-scroll isn't the model — it's the prompt. After thousands of generations on Seedance 2.0, here's what we've learned about writing prompts that actually work.
Think like a director, not a wisher
A weak prompt describes a wish: "an epic dragon." A strong prompt reads like a shot note a director hands to a cinematographer. It answers four questions:
- Subject — who or what is on screen? Be specific: not "a woman," but "a silver-haired violinist in a rain-soaked trench coat."
- Action — what happens over time? Video is motion; a prompt without a verb is a photograph. "She lifts the bow, hesitates, then plays."
- Camera — where does the lens sit, and how does it move? "Slow push-in from a low angle."
- Atmosphere — light, palette, mood. "Neon reflections on wet asphalt, cold teal shadows, warm tungsten highlights."
Put together: "A silver-haired violinist in a rain-soaked trench coat stands under a flickering streetlight. She lifts the bow, hesitates, then plays. Slow push-in from a low angle. Neon reflections on wet asphalt, cold teal shadows, warm highlights on her face."
That's not poetry. It's a work order — and models follow work orders.
Learn ten words of camera language
You don't need film school. These few terms cover 90% of what you'll want:
- Push-in / pull-back — the camera moves toward or away from the subject.
- Pan / tilt — the camera rotates horizontally / vertically from a fixed spot.
- Tracking shot — the camera travels alongside a moving subject.
- Orbit — the camera circles the subject.
- Handheld — subtle shake, documentary energy.
- Static shot — locked-off camera; lets the subject's motion carry the frame.
- Close-up / wide shot — how much of the scene fills the frame.
One camera instruction per prompt is plenty. Stacking three moves in five seconds produces soup.
The mistakes that ruin most prompts
- Overloading. A 5–10 second clip can hold one idea. "A knight fights a dragon while a city burns and refugees flee and…" — pick one.
- No motion. If nothing in your prompt moves, the model invents motion for you, and you probably won't like it.
- Contradictions. "A pitch-black room brightly lit by sunshine" forces the model to average incompatible ideas.
- Style soup. "Cinematic, anime, photorealistic, watercolor" — commit to one look.
Iterate like an engineer: use the seed
Generation involves randomness. The seed pins it down: the same prompt with the same seed reproduces the same result. That turns prompting into a controlled experiment — lock the seed, change one word, and you'll see exactly what that word does. When a generation is almost right, keep the seed and refine the prompt instead of rerolling the dice.
Pick the right mode for the job
Prompts don't work alone — the input mode matters as much:
- Text-to-video when you're starting from pure imagination.
- Image-to-video when the exact look already exists — a product shot, a portrait, an illustration. Your prompt then only needs to describe motion.
- Reference-to-video when a character or product must stay consistent across multiple clips.
Now go generate
Writing prompts is a skill you build in reps, not in theory. Take the four-part anatomy, steal the camera vocabulary, lock a seed, and run five variations of the same idea in the AI video generator. By the fifth clip, you'll be writing shot notes — and the empty prompt box will never look empty again.