This is the first template in the library because it is the least clever one, and the shape it demonstrates is the shape most Wan 3.0 prompts should take.
Read it structurally and there are exactly two parts. Everything before the final full stop is what happens: three sentences of plain narrative prose, no markup, no labels, no numbering. Everything after it is what it looks like: one long comma-separated clause carrying genre, environment, camera behaviour, coverage, particles, colour and mood. That division is not an accident of style. It maps onto how Alibaba's own prompt guidance composes a request — entity, scene and motion first, then aesthetic control and stylization — and it is the reason this reads as a template rather than as one writer's habit.
The thing worth noticing is what is not in it. There is no field protocol, no `description:` heading, no `[Shot 1]` marker, no angle-bracket asset tag and no negative prompt. Wan 3.0 accepts a single plain string up to twenty thousand characters and reads all of it as content. Anything that looks like structure is content too — which is why a prompt copied from a model that publishes a field protocol comes back with the field names rendered into the picture, and why the request succeeds while doing it.
The honest weakness is the beat count. Three things happen: she walks, she discovers, she removes the helmet while birds rise and the sun breaks. That is a thirty-second idea running in five, and it is worth watching exactly how the model absorbs it, because it is not what people predict. It does not drop beats. It cuts: a wide of the ruined avenue, a close-up of the child at the store window with the light in her hands, then the astronaut lifting the helmet off. Three shots in five seconds, one and a bit each. What actually gets dropped is everything the third sentence attaches to that beat — no birds rise, no sunlight breaks through the clouds — because a subordinate clause is the cheapest thing in a prompt for a model to skip.
That is the real lesson and it generalises. Over-writing a short Wan 3.0 request does not buy you a compressed version of your idea; it buys you a montage, assembled by the model, of whichever nouns were load-bearing enough to carry a shot. Verbs in main clauses survive. Atmosphere hung off them in subordinate clauses does not.
Two fixes, and they are different. If you want one shot, cut to one beat: she finds the child, and that is the whole thing. If you want this story, raise the duration to fifteen or thirty and write the order down — "then", "after that", "finally" — because Wan 3.0 has no timecode grammar and connectives are the entire mechanism for saying which thing happens second. An unplanned final third is where a long Wan 3.0 clip drifts, and the word "finally" is what stops it.
The camera instruction is worth arguing with. "Slow cinematic camera movement" gives a speed and no direction and no size, which leaves Wan 3.0 free to pick a move that suits whatever composition it lands on. On a shot this loose that is a reasonable trade. On a product shot it is not, and the fix is the same phrasing the prompt generator on this site emits: the camera pushes in slowly, a small move — a verb, a size and a speed in one clause.
One last thing this prompt gets right by omission: no brand, no on-screen text, no lettering. Ruined convenience stores are covered in signage in every reference image a model has ever seen, and asking for readable text is asking for the one thing generated video still renders badly.
视频