How to write MiniMax H3 prompts
MiniMax H3's six-section prompt format explained, with tags, speaker IDs, dialogue, retention markers, common mistakes and a full before-and-after example.
Last updated 2026-11-05
Black Friday deposit bonus
Every qualifying top-up earns free credit, on top of any other offer.
Ends November 30
MiniMax H3 makes video with sound from reference pictures, but it doesn't read prompts the way most video models do. It expects a structured document in six sections, written in English, with every person and place tagged, every line of dialogue marked up with its speaker and language, and a note on how faithfully each reference must be kept. A one-line request still produces a video. A complete H3 prompt produces the video you asked for: the right person saying the right line, the room from your picture, no surprise soundtrack.
This guide explains the format section by section, lists the mistakes that cost the most retakes, and rewrites a plain request into a full prompt. The source is MiniMax's own prompt writing guide for full-reference mode, which is the mode used when a shot starts from pictures of characters and places. At the end: how the video shots API writes all of this for you.
The six sections
A full-reference prompt is six labelled sections, always in this order:
| Section | What goes in it |
|---|---|
subject_definitions | One line per reference: what it is and what to keep from it |
summary | One short paragraph that starts with the task type in brackets |
retention_analysis | One line per reference: where it appears and how faithfully it is kept |
detailed_description | The video itself, shot by shot, in playback order |
overall_soundscape | Ambience and physical sounds across the whole video |
non_diegetic_music | Score only the audience hears, or N/A |
Everything is in English except dialogue, lyrics and text visible in the scene, which stay in their own language.
The labels: subjects, pictures and audio
References are named with numbered labels, and a label means the same thing in every section once it's defined.
<Subject N>is something visible that the video reuses: a person, an animal, an object, a place, a costume, a style. It's the content, not the file: "<Subject 1> is the young woman in <Picture 1>, with a short black bob and a yellow raincoat."
<Picture N>gets its own line only when the picture is a concrete frame: the shot's first frame, a keyframe, the last frame, or a storyboard. A character sheet or a photo of a place is cited inside the subject's definition instead.
<Audio N>is a sound reference, most often a voice to follow: "<Audio 1> is the voice-timbre reference for <Subject 1> (S1)."
On this API the order is simple: every character and scene you tag becomes a picture, numbered in the order the prompt first mentions them, and a voiced character brings an audio reference.
The summary and its task type
The summary is one paragraph that opens with the task type in square brackets. For shots made from pictures, these are the ones that matter:
| Task type | When |
|---|---|
reference generation | Pictures guide characters, places or style but aren't frames of the video |
keyframe completion | A picture is a real frame: the first, a keyframe or the last |
audio reference | A voice's timbre or delivery is followed without copying the recording |
Combine them with + when several apply: a shot that opens on your picture and has a character speaking in a saved voice is [keyframe completion + reference generation + audio reference]. The summary only uses labels already defined; it never introduces new ones.
Retention: how faithfully to keep each reference
retention_analysis has one line per label, saying where it appears and one fixed marker:
fully_preserved: kept as defined, which is what you want for a character's face, hair and clothes.partially_preserved: still used, with some features changed (the same person in a different jacket).attribute_transfer: a feature moves to someone else (her coat on another character).weak_reference: only a broad likeness of style or mood.
Audio references use their own markers: fully_copy, partially_copy, reference (timbre and delivery only) and weak_reference. A line looks like this: <Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - her short black bob and yellow raincoat are retained.
The detailed description
This is the body of the prompt, and MiniMax asks for detail: for generation, normally 350 to 500 English words. It opens with one or two sentences of style, then runs shot by shot:
[Shot 1]starts the video and has no timestamp. Later shots start with the cut time:[Shot 2] At 00:04.000, the shot cuts to…- At a subject's first clear appearance, describe the referenced features, where it is in the frame and what it's doing. After that, keep using its label.
- Write camera movement in plain words inside the shot: "the camera slowly pushes in".
- Name a frame anchor where it applies: "the shot begins from
<Picture 3>".
Speakers and dialogue#
Every voice gets a stable speaker ID in the order people first speak: (S1), (S2). A subject who speaks carries both labels, and the words go inside <d> with the language:
<Subject 2> (S1) leans forward and says quietly, <d>[English] I thought you weren't coming.</d>Keep the speaker's ID the same every time they speak, end each line with ., ? or ! before </d>, and say plainly when someone stays silent: "<Subject 1> listens and does not speak."
Sound and music
overall_soundscape describes ambience and physical sounds for the whole video: room tone, rain on glass, a coffee machine. Lines of dialogue never go here; they live in the shot where they're spoken.
non_diegetic_music is the score that only the audience hears: instruments, tempo and how it develops. If you don't want music, write N/A. Leaving the section vague is an invitation to add a soundtrack.
Seven mistakes that cost retakes
- Pronouns with two people. "She hands him the cup" is ambiguous when two characters are in the frame. Use the subject label every time a person acts.
- Describing clothes the picture already shows, differently. If the sheet shows a yellow raincoat, "a red jacket" in the description fights the reference. Describe what the picture shows, or leave clothes out.
- A standalone picture line for a character sheet. Pictures that only define a person or place belong inside the subject's line. A separate
<Picture N>line tells the model it's a frame of the video. - Dialogue without its markup. A quote without
<d>[Language] …</d>and a speaker ID may be spoken by the wrong person, or not at all. - Leaving silence implied. If a character shouldn't talk, say so. If there should be no music, write
N/A. - A timestamp on the first shot.
[Shot 1]never has one; every later shot must. - Sections out of order, or in another language. The six sections go in order, in English, every time.
Before and after
Here is a request the way people write it, with two saved characters and a place (pictures 1, 2 and 3, in order of mention):
@Mia sits down across from @Leo in @[Night Café]. She says "You came." He smiles but doesn't answer. Slow push in, no music.And the same shot as a full H3 prompt (shortened in the description for space):
subject_definitions:
<Subject 1> is the young woman in <Picture 1>, with a short black bob, a yellow raincoat over a grey sweater, and silver hoop earrings.
<Subject 2> is the young man in <Picture 2>, with short curly brown hair, a dark green bomber jacket and a white T-shirt.
<Subject 3> is the narrow café at night in <Picture 3>, with a neon sign in a rain-streaked window, small marble tables and warm pendant lights.
summary:
[reference generation] In <Subject 3>, <Subject 1> sits down across from <Subject 2> and speaks one line; <Subject 2> smiles without answering, in a single slow push-in shot.
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - her short black bob, yellow raincoat, grey sweater and hoop earrings are retained.
<Subject 2> (appears in [Shot 1]): fully_preserved - his curly brown hair, dark green bomber jacket and white T-shirt are retained.
<Subject 3> (appears in [Shot 1]): fully_preserved - the neon sign, rain-streaked window, marble tables and pendant lights are retained.
detailed_description:
The target video is realistic and cinematic, with warm low-key interior light and cool neon spill from the window.
[Shot 1] A medium two-shot inside <Subject 3>: rain streaks the window behind a small marble table under warm pendant lights. <Subject 2>, the young man with curly brown hair in a dark green bomber jacket, sits on the right, hands around a coffee cup. <Subject 1>, the young woman with a short black bob in a yellow raincoat, walks in from the left, pulls out the chair opposite him and sits down, raindrops still on her shoulders. The camera slowly pushes in toward the table. <Subject 1> (S1) looks at him and says softly, <d>[English] You came.</d> <Subject 2> holds her gaze, then smiles slowly and does not speak. The push-in ends on both faces in the warm light.
overall_soundscape:
Quiet café room tone, rain against the window, the soft scrape of a chair and a cup set down on marble.
non_diegetic_music:
N/ANotice what changed: every person has a label and a description that matches their picture, the line has a speaker and a language, the silence is stated, the camera move is in the shot, and "no music" became N/A.
Or let us write it
All of this is worth knowing, and none of it is required here. On the video shots API and in the video playground you write the first version, the one-liner with @ tags, and the full prompt is written for you before every render:
Write what you want
Tag your saved characters and scenes with
@, put dialogue in quotes, and say what matters: who stays silent, the camera, no music.A vision model writes the H3 prompt
It reads MiniMax's guide, looks at your actual pictures and writes all six sections: labels, task type, retention, the shot-by-shot description, speakers and sound. Its answer is checked for all six sections in order; a model that fails is retried, and a second model takes over if it keeps failing.
The shot renders
MiniMax H3 gets the complete prompt. The shot object's
prompt_renderedshows exactly what it received, so you can learn the format from your own shots.
The rewrite is on by default and never fails a shot: if no complete prompt comes back, the shot renders from your words the usual way. It runs as an ordinary chat call on your key, shown as other_usd next to the shot's price, and adds 10 to 45 seconds before rendering starts. Shots themselves cost $0.13 per rendered second with one picture.
If you'd rather write prompts yourself, send rewrite: false. A prompt that already starts with subject_definitions: goes to the model exactly as written. The rewrite reference has every field.
Frequently asked questions
Do MiniMax H3 prompts have to be in English?+
The six sections should be in English. Dialogue, lyrics and visible text stay in their own language inside <d>[Language] …</d>, so a character can speak Chinese or Spanish in an English prompt.
How long should an H3 prompt be?+
MiniMax suggests 350 to 500 English words for the detailed description of a generated video. Dialogue-heavy shots may need more, to fit the whole spoken timeline.
What's the difference between reference generation and keyframe completion?+
Reference generation uses pictures as guides for who and where. Keyframe completion uses a picture as an actual frame of the video, such as the first frame. A shot can be both.
Can I see the prompt that was used for my video?+
Yes. When the rewrite ran, the shot's rewritten is true and prompt_rendered holds the full six-section prompt H3 received.
Can I turn off the automatic rewrite?+
Yes, with rewrite: false on the request, or the checkbox in the playground. Prompts that already start with subject_definitions: are always sent as written.
Is this a MiniMax H3 prompt generator?+
In effect, yes: the rewrite is an H3 prompt generator built into the video shots API. You can also use the shots it writes as templates. The model's licence terms are on the MiniMax H3 terms page, and more guides are on the guides page.