Wan Prompts: Structure, Syntax and Examples for Wan 3.0, 2.7 and 2.2

What Wan is and which version is current

Wan is Alibaba's family of video models. You can use it in two quite different ways, and they don't support the same things, so it helps to know which one you're on before you write a word. Checked against the official documentation in November 2026:

• Wan 3.0 (wan3.0-video and the -prime variant): the newest version, available only through the API in Alibaba Cloud Model Studio. Clips of 2 to 30 seconds, up to 1080p, optional audio and reference inputs.
• wan2.7, wan2.6 and wan2.5: earlier versions still offered through the API. All three can generate music or effects that match the scene.
• Wan 2.2: the latest open-weights release, published in July 2025 under the Apache 2.0 licence (T2V-A14B, I2V-A14B and TI2V-5B models). You can run it on your own hardware or through a service that hosts it, but it generates no sound.

In practice the trade-off is simple. The open model gives you control over where and how it runs, without sound or multi-shot. The API versions add audio, voice and timed shots, but you depend on the service and its terms. Neither is better across the board: pick the one that fits the job and check what it supports today.

How a Wan prompt is built

Alibaba's prompt guide works in layers. The basic formula is entity (who or what is on screen), scene (where, and the mood) and motion (what happens). The advanced version adds aesthetic control (shot, camera, lighting) and stylisation (the overall look).

• Text to video: entity + scene + motion + aesthetic control + stylisation.
• Image to video: motion and camera only. The image already sets the look, so there's no need to describe it again.
• With sound (API versions): add a description of the sound at the end.
• Language: the docs list Chinese and English for both the prompt and the negative prompt. English is the safe choice.

Mind the length. wan2.7 accepts prompts of up to 5,000 characters; wan2.6 and wan2.5 cap out at 1,500. Four or five clear sentences are usually plenty.

Wan's own syntax: negatives, shots, references and voice

A few conventions are specific to Wan, and it pays to follow them exactly:

• negative_prompt: a separate API field for what you don't want to see. On wan2.7 it takes up to 500 characters. Write it as a list of terms ("text, logos, watermark, flicker"), not as sentences.
• Restrictions inside the prompt: for sound, short sentences such as "No dialogue." or "No background music." do the job.
• Multi-shot (Wan 3.0, wan2.7 and wan2.6): one overall description, then each shot with its time range: "Shot 1 [0–3 s] …", "Shot 2 [3–6 s] …".
• References: refer to each asset by number, "Image 1", "Image 2" or "Video 1", and say what it's for.
• Voice: the line in quotes, plus emotion, tone, pace, timbre and accent. Wan 3.0 and wan2.7 follow quoted text literally; leave it out and the model makes up what gets said.
• Several characters: give each one a unique label ("the tall bookseller", "the young courier") and stick to it, with no pronouns. Put the action before the line.

On wan2.7 the prompt_extend parameter is on by default, which means a language model rewrites your prompt before generating. If you need your exact wording kept, check that parameter in your tool and turn it off.

Camera and movement

Wan understands plain-English film vocabulary. These are the terms that appear in its documentation:

• Moving closer or away: "camera pushes in", "camera pulls out".
• Sideways moves and turns: "camera moves left", "camera moves right", "panning", "camera tilts up".
• Special moves: "orbiting camera movement", "aerial shot", "tracking shot".
• No movement: "fixed camera".
• Angles: "high-angle", "low-angle", "top-down".
• Shot sizes: "close-up", "establishing shot".

Stick to one camera move per shot and don't pair it with very fast action from the subject. If you need a second move, give it its own shot.

Examples to copy

All characters below are made up, so swap in your own.

A red-haired clockmaker in a leather apron winds a tall brass clock in a cluttered workshop. Warm light from a single desk lamp, dust floating in the air. Close-up, camera pushes in slowly. Cinematic realism, shallow depth of field.

negative_prompt: text, logos, watermark, distorted hands, flicker

The fishing boat rocks gently on the swell while the fisherman in the yellow oilskin coils a rope on deck. Camera pulls out slowly. No dialogue.

A small delivery robot crosses a snowy town square at dusk. Shot 1 [0–3 s] establishing shot, the delivery robot rolls past a lit bakery, camera moves left. Shot 2 [3–6 s] hard cut, close-up of the delivery robot's blinking blue display as snow settles on it, fixed camera. Sound: crunching snow and a distant church bell. No background music.

A tiny bookshop at night, warm lamplight on crowded shelves. The tall bookseller in a grey cardigan places a parcel on the counter. The young courier in a red cap picks up the parcel and says, "Last delivery of the day." Cheerful, quick, bright voice. The tall bookseller smiles and replies, "Then go home and rest." Calm, slow, low voice. Medium shot, fixed camera. No background music.

Aerial shot of terraced rice fields at sunrise, low mist drifting between the terraces. A farmer in a straw hat walks slowly along a narrow path. Soft golden light, documentary look.

negative_prompt: text, logos, buildings, camera shake, flicker

Common mistakes

• Naming real people or asking for a lookalike of someone famous.
• Packing quick scene changes or long choreography into a few seconds.
• Asking for exact on-screen text or word-by-word lip sync, which the official guide itself flags as hard.
• Using "Shot 1 [0–3 s]" on wan2.5 or Wan 2.2, where multi-shot isn't documented.
• Asking Wan 2.2 for sound: it doesn't generate any, so add audio in the edit.
• Writing the negative prompt as "don't show…" sentences instead of a list of terms.
• Forgetting about prompt_extend, then wondering why the video doesn't follow your text to the letter.

Do it faster with PromptNow

PromptNow's Wan prompt generator asks for your version, mode, scene, camera, length and sound, and gives you a prompt to paste into ChatGPT, Claude or Gemini. The AI then writes the final Wan prompt in your version's syntax, with the negative_prompt kept separate, the settings worth checking, two variants and a short explanation of each part.

If the scene itself isn't settled yet, start with the video scene or storyboard template and adapt the result to Wan afterwards.