Seedance 2.5 prompt guide: write prompts that get the shot you pictured

Think of the model as a film crew that has never met you and never asks questions. It reads your prompt once, then shoots. Anything you leave out, it improvises. Write "a woman walking down a street" and the model picks the street, the weather, the wardrobe, the camera angle and the mood for you. Sometimes that is fine. The moment you need something specific - your product, your character, a feeling that matches the rest of your video - improvising is exactly what you do not want.

The fix is not writing longer prompts. It is covering four things in a fixed order, and leaving out the stuff the model cannot use anyway.

Every prompt that works has the same four parts:

• Subject. Who or what the shot is about, described so it can be pictured: "a barista in a grey apron", "a matte black espresso machine on a marble counter". If a clothing detail or a color matters, say it here.
• Action. One clear thing that happens. Pours latte art. Walks through a rainy market. Steam rises from the cup. One action, not five.
• Camera. How the shot is filmed: a slow push-in, an orbit around the subject, a handheld follow, a static frame. More on this below.
• Style and mood. Lighting, color and feel: warm morning light, moody neon, clean commercial look, shallow depth of field.

Stack them into one or two plain sentences and you are done. Here is the formula in action: "A barista in a grey apron pours latte art at a wooden counter, morning light from the left. Slow push-in on the cup, warm and calm, shallow depth of field." Under 40 words, and it covers everything the model needs to stop guessing.

Seedance 2.5 respects camera directions better than most models, but only the ones it knows. These work reliably:

• Slow push-in / dolly in - the camera glides toward the subject. The safest way to add energy to a still scene.
• Orbit - the camera circles the subject. Great for products; ask for a slow orbit or it spins too fast.
• Tracking shot - the camera follows alongside a moving subject. Good for walking and running scenes.
• Handheld follow - the same idea, shakier and more casual. Documentary feel.
• Static, locked-off shot - the camera does not move. Use it when the subject provides the motion, like pouring or dancing.
• Crane up / pull-back reveal - the camera rises or retreats to show the scene. Good closing shots.

One move per clip. A few seconds is not enough time for a pan, a zoom and a crash-in, and asking for all three usually cancels them out. If your video needs several moves, that is a sign you need several shots, not a busier prompt.

This is where most prompt guides go quiet, so here is the honest list:

• Multi-beat action. "He turns, smiles, winks, then walks away" is three shots crammed into one. The model blends them into mush. Split it across clips.
• Quality-word soup. Strings like "8K ultra HD masterpiece best quality" do almost nothing on modern video models. Concrete details - "soft window light", "rain on the pavement" - do the work those words pretend to do.
• On-screen text requests. Asking for a specific sentence to appear in the frame is a coin flip, and text tends to warp. Generate the clean shot, add text in the edit.
• Close-ups of hands doing precise things. Typing, tying, threading a needle. Hands are still where AI video fails most. Frame those moments differently or fix them in post.
• Negative lists. "No blur, no morphing, no extra fingers" mostly gets ignored, and it can pull attention toward the thing you banned. Describe what you want instead.

Five prompts from real jobs, ready to adapt. Swap the subject, keep the structure.

• Product orbit. "A matte black espresso machine on a marble kitchen counter, morning light from the left. The camera slowly orbits around it as steam rises from a cup beside it. Clean, premium commercial look, shallow depth of field."
• Food close-up. "Close-up of a burger on a wooden board, cheese melting slowly down the side. A hand reaches in and lifts the top bun, steam curls upward. Macro lens, warm tones, appetizing commercial look."
• Character scene. "A young courier in a blue uniform jogs up apartment stairs holding a pizza box. Handheld camera follows from behind. Warm evening light through the stairwell windows, energetic and friendly mood."
• App demo shot. "A smartphone lies on a white desk, its screen lighting up with a fitness app. The camera pushes in slowly as a hand picks the phone up. Bright minimal office, crisp daylight, clean tech-commercial style."
• Travel opener. "Aerial shot drifting slowly over a turquoise coastline with white cliffs, a small boat leaving a wake below. Golden hour, sweeping and calm."

Notice none of them are long. Each one fits in two sentences and names the subject, one action, one camera move and a mood. That is the whole trick.

You will hit all of these eventually. Here is what actually helps:

• It ignored my instruction. The important words are buried mid-prompt, or the prompt is doing too much. Put the main action in the first sentence and cut the prompt back to the formula.
• The character looks different every take. Describe hair, clothing and age in the same words every time, or better, use image-to-video with a reference photo so the first frame is locked.
• Everything moves at once and it looks mushy. Too many actions. One per clip, as a rule with no exceptions.
• The clip is boring and dead. Nothing in the prompt moves and the camera is left implicit. Add one subject motion or one camera move.
• Hands, faces or text warp. Reframe the shot to avoid the problem zone, keep motion gentle, and fix small details in the edit instead of rerolling thirty times.

A good prompt gets you one good shot. A real video needs four to eight of them that match - same character, same look, in the right order, with captions on top. That is the gap our studio is built around: plan the shots on a storyboard first, keep character references locked between shots, generate each scene with the right prompt, then trim and caption the cut in the browser. The homepage walks through the full workflow, and the early access list is open if you want to try it.