MiniMax H3 Prompt Master
Pick the right H3 mode and get a timed, sound-directed prompt ready to paste
For creators, marketers and developers using MiniMax H3 (Hailuo 3) — the video model that generates picture and stereo sound in one pass.
Describe your idea in plain words and attach whatever you have: a product photo, a character sheet, two keyframes, a voice clip. The skill:
- Picks the generation mode (text-only, first frame, last frame, first+last frame, or omni-reference) using clear rules — for example, any audio or video input means reference mode, and a character sheet should never be used as a first frame.
- Matches the format to where you generate. Hosted surfaces (Hailuo app, MiniMax Hub, Open Platform API) get an organized Director's Brief; self-hosted open weights (ComfyUI, SGLang) get MiniMax's structured shot/soundscape/music format with correct timestamps, speaker IDs and dialogue tags.
- Plans timing — shots per second, a dialogue word budget, and how to split ideas longer than 15 seconds.
- Writes the audio: ambience, action sounds, voices and score, described the way the model expects.
You get the mode and why, recommended settings (model, duration, ratio, resolution), a copy-ready prompt, optional variants, and a checklist of files to supply with their size and length limits. Limits and syntax were checked against MiniMax's own documentation in October 2026; the skill tells you to trust your live interface if anything has changed since. Results from generative models vary between takes, so the skill also gives you a short iteration ladder for fixing drift.
Highlights
- Chooses T2VA, I2VA, L2VA, FL2VA or Ref2VA from your assets, and says why
- Brief format for Hailuo/API, structured format for local open weights
- Directs dialogue, sound effects and music, not just the picture
- Camera, lighting, sound and music vocab banks plus 5 worked examples
- Settings, asset checklist and a pre-flight self-check every time
Example
H3 prompt please: 8-second vertical video, a cat slowly pushes a glass off a kitchen counter while staring at the camera. Funny. I'll use the Hailuo app.
**Mode:** T2VA — no images or audio, so text-to-video with audio. **Settings:** MiniMax H3 · 8 s · 9:16 · 768p draft, 2K final **Prompt (Director's Brief):** ```text 8s, 9:16. Live-action, bright naturalistic kitchen, shallow depth of field. Cast: a fluffy orange tabby with a white chest and one bent ear. Shot 1 (0–5s): Eye-level medium shot along the counter. The cat sits beside a full glass of water, holds eye contact with the camera, and slowly extends one paw, nudging the glass toward the edge in three small pushes. The camera pushes in slowly with small amplitude. Shot 2 (5–8s): Close-up of the counter edge. The glass tips over and drops out of frame; the cat keeps staring, then blinks once. Sound: quiet kitchen room tone and a fridge hum; glass scraping on stone with each push; a crash and splash off-screen at about 6s; one short meow after the crash. Music: pizzicato strings and a bassoon at a slow, sneaky tempo, stopping dead on the crash. Ending: the cat staring into the lens, completely calm. ``` **You need to supply:** nothing — text only.
I have three photos of our trail sneaker and a 7-second voice memo from our founder. Make a 12s vertical ad where her voice says: "Built for the last mile." We use the MiniMax API.
**Mode:** Ref2VA — the photos define the product but shouldn't be literal frames, and an audio voice reference is only accepted in reference mode. **Settings:** MiniMax-H3 · 12 s · 9:16 (set explicitly) · 2K for the logo and stitching detail Attach in this order → Image 1–3: sneaker (side, sole, top), Audio 1: founder voice memo. **Prompt (Director's Brief):** ```text 12s, 9:16. Live-action outdoor sports commercial, golden-hour side light, warm earthy grade. References: Images 1–3 = the trail sneaker (keep colorway, lacing and sole pattern exactly). Audio 1 = the narrator's voice timbre; she speaks the new line below. Shot 1 (0–4s): Low tracking shot beside a runner's feet on a dusty ridge trail; gravel kicks up with each stride… Dialogue (English, off-screen voiceover, Shot 3): "Built for the last mile." The runner's lips stay closed. … ``` **You need to supply:** 3 product photos (≤30 MB each); the voice memo as WAV/MP3, 2–15 s, one clean voice, no music underneath.
What's inside
minimax-h3-prompt-master/ ├── agents/ │ └── openai.yaml ├── references/ │ ├── anti-patterns.md │ ├── examples.md │ ├── h3-facts.md │ ├── mode-selection.md │ ├── prompt-formats.md │ ├── vocab-camera.md │ └── vocab-light-color-sound.md ├── LICENSE.txt ├── README.md └── SKILL.md
Install by unzipping into your agent's skills folder. Install guide →