Hailuo Prompts: MiniMax-H3 and Hailuo 2.3 Syntax, With Examples

Hailuo, MiniMax, H3: untangling the names

MiniMax is the company. Hailuo is the name of its family of video models, and of its app. MiniMax-H3 is the newest model, often called 'Hailuo 3.0': same model, two names. The earlier models are still available in the MiniMax API as MiniMax-Hailuo-2.3, Hailuo-02 and T2V-01-Director.

Why it matters: H3 and the older models are prompted in completely different ways. Structure, camera instructions and dialogue all change, so check which model you have selected before you write a word.

• MiniMax-H3: text to video; first frame, last frame or both; references (up to 9 images, 3 videos and 3 audio files); and editing. Clips of 4–15 s at 768p or 2K, 24 fps. Prompts up to 7,000 characters. Dialogue in 11 languages, including Spanish.
• Hailuo 2.3, Hailuo-02 and T2V-01-Director: 6 or 10 s clips at 768p, or 6 s at 1080p. Prompts up to 2,000 characters. Camera via bracketed commands. Audio for these versions is not documented.
• Access: the MiniMax API and the Hailuo app. H3's weights are published on Hugging Face under MiniMax's own community licence.

Checked in November 2026. Some sources put H3's release at the end of July 2026. Versions, limits and prices change often, so confirm them in the official docs before you rely on them.

Bracketed camera commands for Hailuo 2.3, 02 and Director

If you are on Hailuo 2.3, Hailuo-02 or T2V-01-Director, write one paragraph in English and request camera moves with bracketed commands. Only these 15 are documented:

• Sideways: [Truck left], [Truck right].
• Pan: [Pan left], [Pan right].
• Towards or away: [Push in], [Pull out].
• Raise or lower the camera: [Pedestal up], [Pedestal down].
• Tilt: [Tilt up], [Tilt down].
• Zoom: [Zoom in], [Zoom out].
• Other: [Shake], [Tracking shot], [Static shot].

Place each command right where the move happens in the text, not all at the start. You can combine up to 3 in one bracket, separated by a comma: [Pan left,Pedestal up]. These versions have prompt_optimizer switched on by default; try your prompt with and without it.

A stone lighthouse stands on a windswept cliff at golden hour, waves breaking white on the rocks below [Static shot]. Seagulls lift off the railing as the lamp flickers on [Pedestal up], then the view turns along the coastline towards the setting sun [Pan right]. Warm low sunlight, long shadows, realistic cinematic look.

A potter in a linen apron shapes a tall clay vase on a spinning wheel in a sunlit workshop [Static shot]. She steps back to look at it as the camera rises and turns at the same time [Pan left,Pedestal up], revealing shelves full of drying bowls. Soft afternoon light through a skylight, documentary look.

How a MiniMax-H3 prompt is structured

H3's official guide splits the prompt into sections, always in this order:

• Instruction line: image modes only. One short sentence saying what to do with the image or references.
• integrated_multimodal_description: everything that is seen and happens (subject, setting, lighting, style, action and camera), with any dialogue woven in.
• overall_soundscape: 1–4 sentences in English on ambience and sound effects. No dialogue here.
• non_diegetic_music: 1–3 sentences on music that isn't heard inside the scene, or N/A.

Camera moves in H3 are plain English with three ingredients: movement type, amplitude and speed, as in 'the camera pushes in with small amplitude at slow speed'. No brackets.

integrated_multimodal_description: A wide shot shows an old ferry crossing a misty lake at dawn, its single lamp glowing on the bow. A ferryman in a wool coat stands at the wheel and slowly turns it to the left. The camera trucks right with medium amplitude at slow speed, keeping the ferry in the centre of the frame. Soft cold light, pale blue palette, realistic cinematic look.
overall_soundscape: Gentle water lapping against the hull, a low diesel engine hum, distant birds waking up.
non_diegetic_music: A quiet solo cello holds long, warm notes.

Dialogue, cuts and image modes in H3

Each speaking character gets an ID in brackets, (S1), (S2) and so on, and their exact line goes inside a <d> tag with the language in square brackets: <d>[English] …</d>. Keep dialogue in the description, never in overall_soundscape.

For more than one shot, leave the first one without a timestamp and open every later shot with 'At 00:03.500, the camera cuts to…' (minutes, seconds, milliseconds). Repeat the character's key traits in each shot so they stay consistent.

integrated_multimodal_description: [Shot 1] A medium shot frames a young watchmaker with round brass glasses at a cluttered wooden workbench, rain streaking the window behind her. She lifts an open pocket watch to her ear; the camera pushes in with small amplitude at slow speed. At 00:05.000, the camera cuts to a close-up of the watch's spinning balance wheel as the watchmaker (S1) says: <d>[English] It's ticking again.</d>
overall_soundscape: Steady rain against glass, soft ticking of many clocks, a faint creak of the wooden stool.
non_diegetic_music: N/A

Use the uploaded image as the first frame.
integrated_multimodal_description: The street vendor in the image starts folding paper lanterns on her stall; steam rises from a pot beside her and a lantern sways in the breeze. The camera pulls out with small amplitude at slow speed to reveal the busy night market around her. Warm lantern light, neon reflections on wet pavement.
overall_soundscape: Crowd chatter, sizzling food, paper rustling, a distant scooter passing.
non_diegetic_music: N/A

MiniMax recommends running H3 prompts through its H3-Context-IR rewriter before generating. Try it, then compare the output with your original prompt.

Common mistakes

• Mixing syntaxes: brackets in an H3 prompt, or sections and <d> tags in a Hailuo 2.3 prompt.
• Inventing commands that don't exist, such as [Orbit] or [Crane up], or packing more than 3 into one bracket.
• Dumping every command at the end instead of placing each one where the move happens.
• Putting dialogue in overall_soundscape, or timestamping the first shot.
• Adding a negatives list: no negative field is documented for these models. Describe what you do want instead.
• Asking for 1080p on H3 or 2K on Hailuo 2.3: each resolution belongs to specific versions.
• Going over length: 2,000 characters on 2.3/02/Director, 7,000 on H3.

Don't ask for the likeness or voice of real people, or for copyrighted characters, brands or music. If you publish the video, label it as AI-generated wherever the platform or the law requires it.

Using the PromptNow Hailuo prompt generator

Rather than memorising every rule, use the PromptNow Hailuo prompt generator. Pick the version (H3, 2.3, 02 or Director), the mode, shot, camera move, duration, resolution and audio, then describe your idea and, if you like, the exact dialogue.

You get a prompt to paste into ChatGPT, Claude or Gemini, which returns the final Hailuo prompt in the right syntax for that version, plus the settings that sit outside the prompt, two variants, a breakdown of each part and a checklist for the tool.