Wan 3.0 Prompting Guide: Write the Shot List
Wan 3.0 puts shots, dialogue, sound and references into a single prompt. Here's how to structure it like a shot list, with example prompts for text, image and reference to video.
A one-line video prompt is easy now. Type "a fox crossing a snowy road at dusk" into almost any model and you'll get something pretty. The trouble starts when the clip needs to run 25 seconds, cut between four shots, give a character two lines of dialogue in a specific voice, and keep the same jacket on the same person from start to finish. That's the job Alibaba's Wan 3.0 was built for, and it changes what a good prompt looks like.
Wan 3.0 renders picture and sound in one pass, runs from 2 to 30 seconds, and reads text, images, video, audio, documents and web pages. Alibaba also published an unusually detailed prompt guide for it. This post turns that guide into a method, adds our own advice where the docs stop, and gives you examples to adapt on each::labs. The short version: write a shot list with a soundtrack.
Pick the Right Wan 3.0 Variant Before You Write a Word
Choose the mode first, because the prompt depends on it. each::labs runs six Wan 3.0 endpoints. Wan 3.0 Text to Video starts from words alone. The Image to Video mode takes a first frame and an optional last frame, while Reference to Video accepts up to 10 reference images, up to 5 reference videos, up to 5 audio clips, plus one file or one public link.
Then there's Wan 3.0 Prime, which Alibaba describes as the high-speed variant with capabilities aligned to the standard model. each::labs runs the same three modes: Prime Text to Video, Prime Image to Video and Prime Reference to Video. Same inputs, same syntax, so everything below applies to both.
The Building Blocks of a Wan 3.0 Prompt
Alibaba's guide is refreshingly direct: one sentence gets you a video, but a more complete and precise description lands closer. Its full formula comes with permission to skip any section you don't need:
[Overall description] + [Reference citation: Image N / Video N / Audio N] + [Shot N (start-end seconds): subject + scene + motion + aesthetic control] + [Dialogue: xx says: "..."] + [Sound effects / BGM] + [Style / Mood] + [Negative list]
The order is the lesson. A one-sentence logline sets theme, perspective and mood. Timed shots follow. Sound and style close it out as global instructions. That's how a director briefs a crew, and it's why long Wan 3.0 prompts stay readable instead of collapsing into adjectives.
Our take: every block is optional except the overall description. A 6-second product spin doesn't need dialogue. A 25-second short drama needs all seven blocks. Prompt length should follow the clip, not your anxiety.
Write Shots, Not Paragraphs
Multi-shot video is where Wan 3.0 rewards structure most. The docs say to number each shot and stamp it with start and end times, end to end with no gaps or overlaps, roughly 2 to 5 seconds each. Brackets or parentheses both work; pick one format and stay with it. Transitions such as "hard cut" or "dissolve" go at the end of a segment.
Here's a 12-second text-to-video prompt we'd start from:
A quiet, observational short about a night-shift baker, warm and tender. Shot 1 [0-4s]: Wide shot of a small bakery kitchen at 4 a.m., flour dust hanging in a single overhead light, the baker kneading dough on a steel table. Static camera. Hard cut. Shot 2 [4-8s]: Close-up, 85mm telephoto, shallow depth of field. Her hands fold the dough, flour lifting off her knuckles. Hard cut. Shot 3 [8-12s]: Medium shot, camera pushes in slowly as she slides a tray into the oven and the orange glow lights her face. Ambient sound only: the hum of the oven, dough slapping the table. No dialogue. No background music.
And when you want the opposite? Write "Generate single shot" or "One continuous shot" on the first line. Alibaba lists that as the way to stop the model from cutting on its own.
Cinematic Video Prompt Language That Wan 3.0 Understands
"Cinematic" on its own tells the model almost nothing. The guide points you to plain camera language instead: push in, pull out, orbit, handheld follow. For a frame that never moves, it suggests spelling it out: "Fixed shot, camera static, position unchanged." Lens and depth cues work too, such as "85mm telephoto, shallow depth of field" or "24mm wide-angle, deep focus", and "High-speed photography (1000fps slow motion)" is the documented way to ask for a slow-motion moment.
Pacing belongs in the overall description: "pacing shifts from steady restraint to explosive release" shapes the whole clip. Our opinion: one camera move per shot. Chaining an orbit, a crane and a push into four seconds is how you get mush. Our camera movement prompts guide has a deeper vocabulary.
Give Your Characters Lines (and Tell Them When to Shut Up)
Wan 3.0 generates dialogue with the picture, so the script lives inside the prompt. The documented format: character, speaking verb, colon, quoted line, as in She whispers: "Don't open it." With several speakers, name who says each line. Voiceover works the same way, and appending "Lip sync" asks for matched mouths.
Voice is where references earn their keep. In Reference to Video, upload a short audio clip and add "Voice timbre reference Audio 1" once; the guide says it then applies to all of that character's dialogue.
A rainy bus stop at night, intimate and melancholic. The woman's appearance references Image 1, her voice timbre references Audio 1. Shot 1 [0-5s]: Medium shot, handheld follow as she steps under the shelter, shaking water off her umbrella. Shot 2 [5-10s]: Close-up, static camera. She looks at her phone and says quietly: "Third time this week." Lip sync. Rain drumming on the shelter roof, a distant bus engine. No background music.
People forget the negative. If you don't want speech, write "No dialogue"; otherwise, Alibaba says, the model decides for itself. A silent product shot that suddenly narrates itself is a wasted run.
Score It or Silence It
Sound in Wan 3.0 works on three layers, and you control each separately. Ambient effects are mostly automatic: describe the action, material and weather clearly (boots splashing through puddles, war banners snapping in wind) and, per the docs, the model fills in matching sound. Specific effects get their own sentences, optionally with timestamps, like "a soft whoosh as the product slides into frame at 3s".
Music gets what the guide calls three-level control. Write nothing and the model picks. Describe it ("cello enters very softly, slow tempo, rhythm in sync with the cuts") and it follows. Or rule it out: "No BGM, generate only ambient and action sounds." Our advice for anything you'll edit later: kill the music, because a baked-in score makes trimming harder. For no audio track at all, switch off the audio toggle on each::labs.
Wan 3.0 Image to Video: Prompt the Motion, Not the Picture
When you supply a first frame, the image already establishes subject, scene and style. Alibaba's guidance is to spend your words on motion, plot, dialogue and camera instead. Redescribing what's in the photo just competes with it.
She lifts the ceramic cup from the saucer, blows across the surface, and takes a slow sip. Steam curls up past her face. She sets the cup down and glances toward the window. Camera static with a very slight natural drift, no cuts. The soft clink of porcelain, quiet café murmur behind her. No dialogue.
Add a last frame and the question changes from "what happens next" to "how do we get there". Describe the path between the two states and let the model fill the middle. One rule to remember: first-and-last-frame mode and reference mode are mutually exclusive, so if you need both a locked opening and a voice reference, you'll have to choose which matters more.
Number Your References Like Evidence
In Reference to Video, materials are numbered by upload order, with images, videos and audio counted separately. The first image is Image 1, the first video is Video 1, and both can exist in the same prompt. Get the order wrong and the model cites the wrong asset. The guide also lets you cite one asset for several jobs, which is the trick most people miss:
Replace the skater in Video 1 with the woman in Image 1, keeping the skating action, trajectory and scene unchanged. Only reference the color tone and lighting of Image 2, not its characters. Upbeat street ambience, wheels on concrete. No dialogue.
Motion from one source, identity from another, grade from a third, all stated plainly. With several characters, lock each to its image and keep their descriptions apart; Alibaba says this stops features bleeding between subjects. Storyboard grids need a warning up front: "Use the storyboard only as shot guidance; do not generate video from the storyboard as a single image."
Files and links are simpler. Wan 3.0 reads the content itself, so the prompt just says who it's for, what style, and what format. The file or page must be publicly reachable; anything behind a login won't load.
Iterate Like an Editor, Not a Gambler
Start small. Our habit: draft at 480P and 5 to 8 seconds, then render the full length at 1080P once the shots behave. Keep a seed fixed while you change one block at a time, though Alibaba notes the same seed won't always repeat exactly.
Know what prompt_extend does before you leave it on. It's enabled by default and uses a language model to rewrite your prompt, which the docs say helps short prompts most. For a timed shot list, test it both ways, since a rewrite can blur your timestamps (with a file or link reference it must stay on). The each::labs schema has no separate negative field, so exclusions go at the end of the prompt. The official Wan 3.0 prompt guide also offers a prompt-optimization skill called wan3-pe, plus many full-length examples.
Mistakes That Waste a Wan 3.0 Run
The biggest one is padding the negative list. The guide says to list only what you don't want and never repeat the positive prompt; twenty variations of "no blur" just dilute what matters. Close behind is asking for high-frequency actions the model can't control. Alibaba's own example is a head that shakes three times every second. Describe the gesture, don't count it.
Leaving audio unstated is the third. Silence is a choice you have to write down. The fourth is putting settings in prose: ratio, resolution and duration have their own fields on each::labs. Our Wan API parameters breakdown covers those fields in detail.
Wrapping Up
With Wan 3.0, one prompt carries a whole production: shots, camera, cast, script, sound and grade. Open with a logline, time your shots, number your references, quote your dialogue and write down what should stay silent. Then iterate one block at a time. Browse the wider text-to-video catalog when you want a second opinion from another model, and borrow the conversational habits from our Gemini Omni prompting guide where they fit.
Frequently Asked Questions
Is Wan 3.0 Prime prompted differently from Wan 3.0?
No changes needed. Alibaba positions Prime as the high-speed variant with capabilities aligned to the standard model, and on each::labs both expose the same text, image and reference modes with the same inputs. Write once, then pick whichever fits your turnaround.
How long should a Wan 3.0 prompt be?
As long as the clip demands. The model accepts up to 20,000 characters, but a 5-second shot rarely needs more than three sentences. Multi-shot clips with dialogue need more, and the structure (logline, timed shots, audio, style) matters more than the word count.
What's the best way to use Wan 3.0 image to video for product shots?
Let the photo do the describing. Spend the prompt on the motion, one camera move and the sound, then add "no cuts" and "No dialogue" for a clean hero shot. If the end state matters, supply a last frame and describe the path between the two.
Can Wan 3.0 generate a full 30-second clip in one go?
Yes, duration runs from 2 to 30 seconds, and each::labs also offers an auto setting that lets the model choose. Draft shorter first, though. Long runs carry more risk of drift and stray cuts, so lock the shots at 5 to 8 seconds before you commit to the full length.