MiniMax H3 Official Handbook: A Three-Part Formula That Turns Prompts from Scene Descriptions into Asset Instructions

First, know what H3 can take in and put out

Before writing a single word, know the model's capability envelope — it dictates how you write the prompt. The 7,000-character limit, for instance, explains why official examples routinely run thousands of characters — MiniMax genuinely wants you to fill the space.

On the input side, there are two very different entry points. Pick the wrong one and the best prompt in the world won't save you:

The three capability pillars the handbook defines for H3

Knowing what the model is actually good at tells you where to point your prompt. The handbook opens with an official categorization, quoted directly:

The first two pillars each break down into four sub-capabilities, and the handbook's examples are organized along those lines:

Part one: assign a job to every asset

This is the biggest mental shift in the whole handbook. In the past, the first thing you did was describe the picture. Now the first thing you do is go through every asset you uploaded and say what it's for.

Two things to cover: the label (in upload order, @image1, @audio2, @video3) and the job description.

MiniMax lists the "jobs" as a vocabulary of 13 labels. The list itself is copy-paste material; scan it when writing part one:

@image1 provides the appearance of character XX (if there are multiple characters, make sure the reference is unambiguous), @video2 provides the action reference.

Two practical extras. If a specific trait in your reference really must survive, say so explicitly — consistency improves. For voice cloning, if the dialogue or lyrics have to stay intact, MiniMax strongly recommends pasting the full lyrics into the prompt, e.g., "@audio1 serves as the voice-cloning source for the target video, with the exact lyrics: 'ABCDEFGHIJKLMN'."

In practice, the "voice reference" label works like this: give an audio clip as the timbre reference, give it a line to speak, and the character delivers that line in that voice.

The last label, "video editing," covers modifying an existing clip. The simplest prompt is plain language: "replace the cat in the video with a dog."

If you didn't upload anything, skip this section.

Part two: one sentence to frame the whole video

With asset assignments done, establish what this video is. The core creative concept should lock the piece in one sentence. MiniMax requires four things in it: subject (who/what), location (where), event (what's happening), and genre/style (live-action / animation / cinematic / commercial...), plus an optional special camera move (aerial / one continuous shot / slow motion).

A young woman in Hanfu (@image1) practices swordplay in a courtyard where cherry blossoms drift through the air — classical Chinese aesthetic, cinematic, one continuous shot.

Two hard rules hide in this section, rarely mentioned elsewhere.

MiniMax also lists specific options for "genre/style" and "cutting style" that you can pick from directly:

Genre/style: realistic / animation / cinematic / ad / documentary...
Specific styles: cyberpunk / neon aesthetics / graffiti style, etc.

Special camera moves: aerial / one continuous shot / slow motion...
  · Cuts happen by default
  · Camera style: be specific. For orbit shots, use truck left + pan right
    or truck right + pan left rather than saying "orbit shot"
  · Cutting style:
    · Standard cuts
    · Dissolves (fade)
    · Cuts synced to the rhythm
    · Quick cuts

If any element links directly to a reference asset, feel free to emphasize it with @image1 / @audio1 / @video1.

One more text-related rule: if you want specific text, a logo, a headline, or button copy to appear in the video, you must write the exact string into the prompt. For example, "The phone screen displays the title: 'AI Video Creation' with the button text: 'Start Now'." In the handbook's full-of-text motion examples, every character is hard-coded into the prompt:

Part three: write what you want — and don't want — on a timeline

Once the overall tone is locked, you finally write what happens in every second. This section is organized into timeline blocks, with each cut acting as a timestamp, and each block covering six items.

0–3s: Wide shot. The woman (@image1) walks slowly into the cherry-blossom courtyard (@image2) from screen left, background blurred, no dialogue, just footsteps.
3–8s: Cut to a medium shot of the woman (@image1). She draws her sword (@image3), slowly assumes a stance, and cherry petals fall from the tree. Camera pushes in.
8–12s: Cut to a close-up. A flash of the blade, slow motion, petals scatter from the force of the sword.
Non-diegetic music: N/A

That last line is how you write "don't want." If you want no background music, saying "no BGM" doesn't work. You have to write this:

Non-diegetic music: N/A

Or in English format:
"non_diegetic_music": N/A

Two more notes: favor concrete, visible images over metaphors — H3 is built for explicit visual instructions, not abstract similes; and if there's dialogue, write the exact lines, don't just say "she says something cutting."

What does this section look like at full power? One handbook prompt asks for four simultaneous changes, and every one is specified to the letter:

If dialogue and shots are out of sync, the lip-sync breaks

But there's a storyboard trap the handbook calls out specifically. It adds a note after this rule: "This one needs stressing — a lot of lip-sync problems come from exactly this."

Three related rules:

If dialogue spans multiple shots, say which shots it spans. H3 can handle J-cuts and L-cuts, but you have to spell them out.

Make clear who is speaking and whether they're on screen. The handbook gives a complex example:

A voice-over begins: Wake up, wake up. Then it cuts to a medium close-up of a middle-aged woman, owner of the voice, who continues: It's time to go to school!

When you cut, say what shot size you're cutting to and which earlier character is the subject — cross-shot consistency stays much better.

Three modes, three prompt-writing styles

The same formula changes depending on how you feed it. The handbook splits it into three modes:

That middle row is the trap: a lot of people send first/last frames hoping H3 will "perform" between them, and it just obediently fills in the motion with no shot changes. That's by design, not a failure.

Six most common pitfalls, as listed by MiniMax itself

With the writing rules covered, here's the handbook's own summary of mistakes. This table belongs next to your keyboard:

Two thousand-character full examples you can copy

Rules only get you so far. At the end of the handbook, MiniMax provides two fully written examples — complete prompts over a thousand characters each, strictly following the three-part structure.

Example one: the on-screen-lyrics MV

Trap music video aesthetic. The most instructive part is a dedicated "groove rules" section that binds the beat to specific on-screen actions:

hi-hat roll → rapid micro-shakes, frame skips, text splitting into fragments
snare → text suddenly enlarges, hard cut, character's shoulders press down
808 bass hit → low-end presses the frame, brief image distortion, text stretches vertically or compresses horizontally
vocal keyword → lip-sync, jaw movement, head-bobbing, hand gestures push forward

Example two: hand-drawn effects

A live-action tram car fused with hand-drawn glowing animation. The most instructive part is the shape-continuity requirement: one hand-drawn line changes form a dozen times across 15 seconds (ticket → paper swallow → caterpillar → arrow → little sailboat → mini tram → snail → umbrella → small fish → sunset cloud sea), and MiniMax requires each transformation to retain traces of the previous form — the ticket's dashed border, the swallow's wings, the caterpillar's dots — so the audience feels it's one thing continuously transforming, not a new character appearing.

It also specifies the camera's rhythm: the camera always trails the hand-drawn animation by half a beat, letting the animation run first and the camera hesitantly chase after it.