How to prompt Hailuo: fifteen camera commands, written in square brackets

Hailuo is MiniMax's picture-only video line, and it directs the camera in a way no other family in the LUVI catalog does: with a bracketed command taken from a closed list of fifteen, dropped inline where the move happens. On LUVI it runs as Hailuo-02 and Hailuo-2.3, each in a Fast, Standard and Pro tier, from text or from an image. On September 22, 2026 a six-second clip quoted 380 credits on Hailuo-2.3 Fast, 560 on Standard and 980 on Pro, with the older Hailuo-02 Fast down at 200. Here is how the model reads a prompt, starting with those brackets.

What are the square brackets in a Hailuo prompt?

They are camera commands, and MiniMax publishes the full list: fifteen of them, in exact spelling. The documentation gives the shape with its own example, A man picks up a book [Pedestal up], then reads [Static shot].

• Truck: [Truck left], [Truck right]
• Pan: [Pan left], [Pan right]
• Push: [Push in], [Pull out]
• Pedestal: [Pedestal up], [Pedestal down]
• Tilt: [Tilt up], [Tilt down]
• Zoom: [Zoom in], [Zoom out]
• Shake: [Shake]
• Follow: [Tracking shot]
• Static: [Static shot]

That list is closed. [Dolly], [Orbit], [Crane] and [Rack focus] are not commands, and inventing one does not fail loudly—it simply does nothing that the bracket promised. MiniMax is direct about the alternative: "Free-form descriptions also work, but explicit commands yield more accurate results." So an unlisted move belongs in plain prose, where it binds less precisely, and a listed move belongs in its bracket.

How do I combine or sequence camera moves?

By deciding whether the moves happen at once or one after the other, and writing them differently. Commands inside a single pair of brackets run simultaneously—MiniMax's example is [Pan left,Pedestal up]—and the documentation recommends a maximum of three. Moves that follow each other go in separate brackets, in the order the text reads, as in "...[Push in], then...[Push out]".

One clip carries one subject, one primary action, one setting and one camera treatment. Hailuo's strength is physical plausibility, so describe motion as motion—what moves, in which direction, how fast, what it touches—and drop the praise words. "Beautiful", "stunning", "masterpiece" and "8K" buy nothing here.

A barista tamps ground coffee into a portafilter on a scratched steel counter, then knocks the basket against the bin. Warm morning light from a window on the left, steam drifting through it. [Push in], then [Static shot]. Realistic film.

That example is written for ten seconds, because it carries two beats and two moves. A six-second clip holds one of each: cut the second half rather than speeding both up.

Why is there nothing to say about sound?

Because Hailuo does not make any. The line renders picture only: no dialogue, no sound design, no music. This is worth stating plainly, because most of the video families next to it in the catalog do generate audio in the same pass, and the habits carry over. Every sentence you write about room tone or a spoken line in a Hailuo prompt is a sentence the model spends on nothing. The help center keeps the current list of which models generate sound with the video.

There is no negative prompt on this line either, and no seed. Exclusions are written as the positive state you want: "an empty platform" rather than "no people", "sharp focus throughout" rather than "no blur".

Should I leave the prompt optimizer on?

That depends on whether you want the model's phrasing or your own. The parameter appears in LUVI as enable_prompt_expansion and it is on by default; MiniMax's documentation describes the same switch and says to "set to false for more precise control". Left on, it rewrites your prompt before the render and also runs the safety checker. Turn it off when you have already written the shot exactly—especially when you have placed brackets and don't want them reworded.

The prompt itself is capped at 2,000 characters, and the sweet spot is well under that: one clear paragraph.

Which Hailuo model, and how long can the clip be?

Three tiers, and the duration depends on which one you pick. On LUVI, Hailuo-2.3 Standard is the tier that lets you choose between six and ten seconds; the Pro tier has no duration control at all and renders six. That mirrors MiniMax's own matrix, where the higher resolution is available at six seconds only. Ten seconds on Standard quoted 1,120 credits against 560 for six, so length scales the price directly.

For image-to-video the rule changes shape. The supplied image is the first frame, so subject, wardrobe, composition and light are already fixed; the prompt should carry motion and a command or two, nothing else. MiniMax's own example is the entire prompt: "A mouse runs toward the camera, smiling and blinking." Re-describing the picture wastes the words, and asking for a gesture the frame cannot support produces a broken one.

How we checked

• Models: Hailuo-2.3 t2v Standard and t2v Pro, with the Fast and Hailuo-02 tiers read from the catalog.
• Settings: six-second and ten-second clips from text.
• Date and what was read: September 22, 2026—the models' parameter contracts and live credit quotes in LUVI, LUVI's own prompting guide for the Hailuo family, and MiniMax's text-to-video API reference.
• Results: 560 credits for six seconds and 1,120 for ten on Hailuo-2.3 Standard; 980 for Pro; 380 for Hailuo-2.3 Fast and 200 for Hailuo-02 Fast. No outputs were generated for this post, so it makes no claim about output quality.
• Drawbacks: the recommendation to write an unlisted move as prose follows from MiniMax's own wording, not from a published comparison of the two forms.

In LUVI

Open a Hailuo model from the Hailuo family page, write one paragraph, and put the brackets where the move belongs rather than at the end. The feather button beside the prompt box applies LUVI's prompting guide for the selected model, which knows the command list and will not invent a sixteenth. If you are moving here from a model that generates its own audio, check when a negative prompt exists too—the habits from those families don't transfer.

New accounts start with 1,000 credits, which covers a couple of six-second clips, so you can create a LUVI account and try the bracket syntax before deciding whether this is your family.

Sources

• Video generation (text-to-video) API reference — MiniMax, official API documentation. Accessed September 22, 2026.