Every other long template on this page says the same thing: past fifteen seconds, name your beats or the model invents them. This one is the exception, and understanding why is more useful than the rule on its own.
The rule exists because unassigned time gets filled. A prompt with three beats in a fifteen-second request leaves a third of the clip unspecified, and the model resolves the ambiguity with a drift and a fade. But "unassigned" is about instructions, not about events. A continuously travelling camera moving through a continuously changing environment is fully specified at every moment even though nothing happens: the instruction is "keep going through this", and it applies to second fourteen exactly as much as to second two.
So the useful distinction is not short prompt versus long prompt. It is whether the prompt describes a sequence of states or a single continuous process. Sequences need connectives and an ending. Processes need a direction, a texture and an exit.
This prompt has all three. The direction is forward and low, down the canyon. The texture is the lens work — frost on the element, barrel distortion, a high shutter that separates the powder into grains. The exit is the burst of light closing over the frame, which is what turns fifteen seconds of travel into a shot with a shape. Take the exit away and you have footage rather than a clip, and the model will supply its own ending, which for a travel shot is usually a slow fade in the middle of the move.
The two changes in the middle — the creature rearing up, the drop into open snow — are worth looking at closely, because they are not beats even though they are written with "then". Neither is something the camera stops for or a new situation the shot has to establish. They are things that happen to a camera that keeps doing what it was already doing. That is the shape of a change that can live inside a continuous shot without fragmenting it, and it is a distinction worth being able to make when you are deciding whether your fifteen-second idea needs four sentences or one.
The look clause is the other thing to take away, and it is almost entirely made of lens facts rather than adjectives. "Frost on the front element" and "slight barrel distortion" assert that a physical object was present in the cold. Neither is beautiful; both are the kind of imperfection a real camera produces and a rendered image does not think to include. Asking for the artefacts of a real recording is consistently more effective than asking for realism, and it costs about eight words.
One last note on the audio, because this subject demonstrates the point better than any other. Wan 3.0 writes sound in the same pass as picture, and it has strong priors about what kind of sound belongs with what kind of picture. Fantasy action has been scored and narrated in nearly every example the model has ever seen, so an unwritten sound layer here does not return silence — it returns a soundtrack and, frequently, a voice. The last sentence of this prompt refuses both, separately, because they are two different habits and refusing one does not refuse the other. As with the rest of this group, the clip arrived without a prompt and this one was reconstructed from the frames.
视频