01
Describing a photo instead of a video
Beautiful scene, zero verbs — the model has nothing to animate, so it invents drift. Fix: every shot gets one observable change, stated in playback order.
MiniMax H3 prompting guide
The MiniMax H3 prompt guide: direct, don’t describe.
MiniMax H3 turns one prompt into a finished shot — picture, motion, dialogue and stereo audio together. This guide covers the prompting rules the model actually rewards, and the free Prompt Agent applies every one of them for you in a short chat.
The shortcut
Meet the Prompt Agent: a director who interviews you.
Most ideas don’t fail as ideas — they fail in translation to camera moves, timing and audio fields. The Prompt Agent closes that gap: describe your video in one plain sentence and it hands back a production-ready MiniMax H3 prompt.
Chat is free
Agent conversations never consume credits. Credits are only used after you continue to the generator and confirm the render — with the cost shown first.
One question, not twenty
The Agent asks only when the answer would change the result. Everything else — camera, lighting, sound, pacing — it decides with professional defaults.
Understands your uploads
Attach photos, video or audio. Each file is assigned one clear job in the prompt, and the Agent routes you to the right mode: text, frames or full reference.
16 creative templates
Start from a cinematic trailer, UGC product demo, talking head, beat edit or twelve more proven structures instead of a blank box.
From idea to prompt in three steps
Open the Prompt Agent
A free account keeps your briefs, prompts and renders together.
01
Describe the video in one sentence
Plain language is enough — no cinematography vocabulary required. Attach references if you have them.
02
Answer at most one question
If something would genuinely change the shot — who the subject is, what a reference is for — the Agent asks once, with quick options.
03
Continue to the generator
The finished prompt, mode, model and duration carry over in one click, with the credit estimate shown before anything is spent.
The formula
Eight elements every strong H3 prompt covers
You won’t need all eight every time. But when a render disappoints, the missing element is almost always on this list.
01
Subject
Name who or what the shot is about — one main subject with a concrete identity. “A gray-haired watchmaker in a leather apron” gives the model something to hold; “a person” does not.
02
Action
Give the subject an observable change: what they do, what happens to them, what looks different by the final second. A prompt without change describes a photo, not a video.
03
Environment
Anchor the scene in a place the camera can show — surfaces, weather, time of day, background life. Environment is what makes motion read as real.
04
Camera
One move per shot, said in camera language: push in, pan, truck, tracking, static — with amplitude and speed when they matter. “The camera pushes in with small amplitude at slow speed.”
05
Lighting
Say where light comes from and how it behaves: “a single warm desk lamp, shadows swinging as it rocks.” Light direction does more than a stack of mood adjectives.
06
Style
Set the look once — film stock, animation style, color palette — at the start of the prompt, instead of sprinkling adjectives through every sentence.
07
Sound
H3 generates stereo audio natively. Keep dialogue with its speaker, put ambience and physical sounds in their own field, and only request background music you actually want.
08
Ending
Decide the final frame — a payoff, a reveal, a held composition. Prompts that end nowhere produce videos that end nowhere.
Failure modes
Five mistakes that ruin H3 renders
Every one of these comes from real prompts. The fix is always cheaper than a re-render.
02
Stacking camera moves
A push-in that also pans while orbiting reads dramatic on paper and renders as jitter. Fix: one camera instruction per shot; change shots when you need a new move.
03
Leaving choices in the prompt
“A red or blue dress”, “something like…” — the model picks for you, randomly. Fix: decide before you render. A final prompt contains decisions, not options.
04
Ignoring the audio fields
H3 ships sound whether you direct it or not — undirected audio means a random soundtrack over your shot. Fix: separate dialogue, ambience and music; write N/A for silence.
05
One shot doing five shots’ work
A full story arc crammed into a few seconds renders as a blur of half-finished beats. Fix: budget actions to the duration, and only cut when the cut reveals new information.
Worked example
From a one-line idea to a directed shot
What most people type
A cool cinematic video of a barista making latte art, beautiful lighting, high quality, 4k, trending
The directed version
integrated_multimodal_description: [Shot 1] Warm 35mm film look. A close-up frames a barista’s hands tilting a steel pitcher over an espresso cup; white foam blooms into a rosetta as she finishes the pour. The camera pushes in with small amplitude at slow speed. [Shot 2] At 00:05.000, cut to a static counter-level shot: she slides the cup toward the lens and the rosetta settles. overall_soundscape: Espresso machine hiss fading out, milk swirling against steel, ceramic sliding on wood. non_diegetic_music: N/A
- Every sentence is filmable — an action, a camera move or a sound, never a quality claim.
- Each shot carries exactly one camera instruction, and the cut lands on a timestamp.
- Picture, physical sound and music are separated, and unwanted music is declared N/A.
This is the rewrite the Prompt Agent performs on your idea — free, in chat, in under a minute.
Rewrite my ideaMiniMax H3 prompting, answered
How do I write a good MiniMax H3 prompt?
Write it like a directing brief, not a caption: one concrete subject, an observable action, a real environment, one camera move per shot, and explicit audio. If you’d rather not learn the craft, the free Prompt Agent builds the brief from a plain-language sentence.
Is there an official MiniMax H3 prompt format?
MiniMax’s guide splits a prompt into three fields: integrated_multimodal_description for the visual timeline, overall_soundscape for physical sound, and non_diegetic_music for score. Short natural-language prompts also work — use the structure when timing, cuts or audio control matter.
Do I need to know cinematography terms?
No. Camera vocabulary makes results more predictable, but the Prompt Agent exists precisely so you don’t have to learn it — it converts plain language into shot scales, camera moves and audio fields.
Is the Prompt Agent really free?
Yes — chatting with the Agent costs nothing and never touches your credits. Credits are only spent when you continue to the generator and confirm a render, and the estimate is shown before you do.
Where can I see real H3 prompts with their videos?
Our verified prompt library pairs MiniMax-published H3 videos with the exact prompts behind them, each linked to its primary source — the fastest way to calibrate what a given prompt style produces.
Stop guessing. Interview your way to a better prompt.
Open the Prompt Agent, describe the shot in your own words, and walk into the generator with a production-ready MiniMax H3 prompt.