Seedance AI: The Prompt Structure That Fixed My Hit Rate (2026)

Seedance is the model I actually render with, most days, for two brands that pay their own way. I've put several hundred prompts through it. For the first month my hit rate was terrible — maybe one usable clip in five — and the fix turned out not to be a better tool but a better prompt shape.

This is what Seedance AI does well, where it reliably breaks, and the prompt structure that took me from one-in-five to roughly three-in-four.

What Seedance actually is

Seedance is a generative video model: you give it a text prompt, usually an image to start from, and it produces a short clip with native audio. It sits in the third category of AI video makers — not a template assembler, not an avatar reader, but a model that builds footage that never existed.

Two properties matter more than any spec sheet. First, it holds identity well when you start from an image — the same face survives a clip far better than it does with text-only generation. Second, it generates sound with the picture, which removes a whole editing step for the kind of short, native-looking clips that actually perform.

What you trade for that is variance. Seedance rewards precise prompts and punishes vague ones harder than the template tools do, which is exactly why prompt structure is the whole game.

The prompt structure that fixed my hit rate

My early prompts were one long sentence describing a vibe. That produces the model's average, and the average is a perfume commercial. What works is naming four things explicitly, in this order.

1. The subject and the frame. Who is in shot, how much of them you see, what they are doing. "A woman in her forties, waist-up, pulling on a jacket" — not "a stylish woman".
2. The camera. This is the highest-leverage line in the prompt. "Handheld phone camera, slightly uneven framing" and "50mm, shallow depth of field, slow push-in" produce two completely different genres from the same subject.
3. The light. Name it. "Overcast window light from the left", "warm late-afternoon sun", "ordinary indoor ceiling light". Left unspecified, the model defaults to dramatic, and dramatic reads as advertising.
4. The motion, timestamped. Describe what changes across the clip rather than a static tableau: "0–2s she reaches for the zip, 2–5s she turns toward the window". Static descriptions produce clips where nothing happens for five seconds.

Then a short exclusion line — "no text, no logos, no watermarks, no music" — because on-screen text is where the model is weakest and a rendered logo ruins an otherwise usable take.

Start frames beat text prompts

The single biggest quality jump was to stop generating from text alone. Generate a still first, approve the face and the garment, then animate from that frame.

Two reasons. You can iterate on a still for a fraction of the cost of a clip, so you approve the expensive part cheaply. And identity drift — the face subtly changing shape mid-clip — mostly disappears, because the model is interpolating from something fixed rather than inventing from scratch. The broader technique is covered in the guide to consistent AI video characters.

Practically: I now spend maybe four still generations getting the frame right, then one or two clip generations. That is dramatically cheaper than six clip generations, and the result is more controlled.

Generate short, then stitch

Ask for five to eight seconds per shot, never fifteen in one go. Long generations drift — the room changes, the jacket changes colour, the face softens — and a drifted fifteen-second clip is a total write-off. A drifted six-second shot costs you one cheap retry.

It also matches how the finished thing should be cut. The ad structure that works for me is a hook shot, a proof shot and a close, which is three short generations stitched, not one long take. Cutting between shots every four or five seconds is native grammar on a phone feed anyway.

What it's still bad at

• Hands doing precise things. Fastening a clasp, holding a small object, typing. Cut away before the hand does anything complicated.
• Any on-screen text. Prices, labels, packaging copy. Exclude text in the prompt and burn it on afterwards where you control it.
• Crowds and head-counts. Ask for three people and you get somewhere between two and five. If the count matters, state it as a hard number and expect to retry.
• Continuity between separate clips. Two generations from the same prompt are siblings, not the same take. Use the same start frame if the shots must match.
• Reading your intent. It optimises for the words you wrote, not the ad you imagined. Every failure I've traced back was an under-specified prompt, not a model limitation.

Does the output actually convert?

Worth separating two questions: does it look good, and does it sell. In my testing those pull in opposite directions.

The most beautiful Seedance clip I've made — controlled lighting, slow dolly, genuinely nice — returned 0.9 on ad spend. A deliberately scrappier one, prompted for handheld phone framing and ordinary indoor light, returned 3.3 on the same product, same budget, same day.

The organic numbers agreed. The clip that opened with a specific surprising number pulled 796 views on a channel where the surrounding uploads sat between 1 and 6. Nothing about the render was better; the first two seconds were.

Which is the practical conclusion: use Seedance's quality to make something believable, not to make something impressive. Prompt for the imperfection. Then put all remaining effort into the opening line — see the ad structure guide for the fifteen-second skeleton I use.

FAQ

What is Seedance AI?

A generative video model that produces short clips with native audio from a text prompt and, optionally, a starting image. It's used for short-form social video, product clips and ads rather than long-form editing.

Is Seedance better with an image or text alone?

With an image. Generating a still first, approving it, and animating from that frame gives far better identity consistency and costs less overall than iterating on full clips.

How long should a Seedance clip be?

Five to eight seconds per generation. Stitch two or three for a fifteen-second ad — long single takes drift and are expensive to retry.

Why do my Seedance clips look like adverts?

Because the prompt asked for cinematic language. Name a handheld phone camera and ordinary indoor light instead, and exclude music. The model will happily produce ordinary if you request it.

Can Seedance render text or logos?

Not reliably. Exclude text in the prompt and add any wording afterwards in an editor, where you control spelling and placement.

How many attempts does a usable clip take?

With a vague prompt, about one in five. With subject, camera, light and timestamped motion all specified — and a start frame — closer to three in four.

Where to go next

Take a prompt that failed for you and rewrite it as four lines: subject and frame, camera, light, timestamped motion. Add one exclusion line. Generate a still first, approve it, then animate. That single change moved my hit rate more than any setting I've touched since.

If you'd rather work through it alongside people running the same tests, I run a free Skool community where we post the prompts and numbers that actually landed — the failures too. Come say hi.

Or, if you'd rather hand the whole thing over, that's something I offer as a service.