MiniMax H3 prompt skill: the exact format the model was trained on

TL;DR

• The skill turns an idea, an image or a clip reference into a prompt in MiniMax H3's trained format instead of free-form prose — the README calls the format "the single biggest lever on output quality" for the model.
• It picks one of five modes first: T2VA (text only), I2VA (image as first frame), FL2VA (first and last frame), L2VA (image as last frame) and Ref2VA (loose style, motion or voice references).
• It copies the format verbatim from references/base-en.txt and references/ref-en.txt: numbered reference jobs, [Shot N] At MM:SS.mmm timestamps, camera moves with type plus amplitude plus speed, dialogue inside <d>[Language] …</d>, and the two mandatory soundscape fields.
• Duration and aspect rules are documented per interface — direct API, local frame grid, or a provider — because they do not transfer between surfaces.

What it is and who it is for

MiniMax H3 Prompt Writer is an agent skill for MiniMax H3, the omni-modal video model that generates synchronized native audio — dialogue, sound effects and music — from structured English prompts. H3 reads those prompts as a contract, so a prompt written as loose prose wastes the model.

The skill is for the person already generating H3 clips who keeps getting near-misses: the wrong cut count, dialogue that drifts, audio that ignores the scene. Rather than another list of tips, it writes the prompt the model was trained to read.

What it does

The README states five jobs:

• Route the input to the right mode. T2VA for text only, I2VA when the image is the first frame, FL2VA for first and last frame, L2VA when the image is the last frame, Ref2VA when the images are loose style, motion or voice references. It asks one routing question only when the mode is genuinely ambiguous.
• Copy the trained format verbatim. Numbered reference jobs, [Shot N] At MM:SS.mmm timestamps, camera moves described by type, amplitude and speed, dialogue in <d>[Language] …</d> with stable speaker IDs, on-screen text in quotes, and the two mandatory soundscape fields.
• Get duration and aspect right per interface. Through the direct API, T2VA needs a concrete ratio while I2VA follows the image and ignores a conflicting one. Locally, the frame grid is 17*n+5 at 24fps. The README is explicit that there is no schema transfer between surfaces, and it covers segment budgets plus the two-decimal alignment that FL2VA and L2VA need.
• Self-validate silently. A mechanical checklist runs before output — alignment line, field order, timestamps, closed tags, quoted text, length budget — and the user only sees the finished prompt.
• Repair failures locally. Paste what went wrong and it loads references/troubleshooting.md, diagnoses symptom to cause, and emits a repair prompt that reuses the good clip as a reference and names the defect precisely.

How to use it

Copy the folder into your agent's skills directory:

git clone https://github.com/artificialguybr/minimax-h3-prompt-skill.git
cp -R minimax-h3-prompt-skill ~/.claude/skills/

SKILL.md loads on demand and the reference files load lazily, so keep the relative paths intact. Then ask for an H3 prompt in the shape you need: a scene with duration, aspect ratio and dialogue language; an opening and closing frame for FL2VA; or a repair for a clip with a specific defect.

The output is the formatted prompt only — no commentary, no code fences. Environment quirks are built in, such as CFG=1 making negative prompts useless and the low-resolution-first workflow.

What is inside the skill

SKILL.md is a compact router and the reference files load lazily: capabilities and input roles, the trained format for T2VA/I2VA/FL2VA/L2VA in base-en.txt and for Ref2VA in ref-en.txt, shot direction, audio and voice, the hosted freeform format, editing briefs, examples, a symptom-to-cause troubleshooting file, an evidence ledger of 18 primary sources at three confidence levels, and community findings marked experimental. A scripts/lint_prompt.py linter checks a prompt mechanically.

When to use it, when not to

Use it whenever a prompt has to survive contact with H3: multi-shot scenes, dialogue with a specific language, sound effects tied to an action, or a repair run after a bad take. The structured format is what keeps a five-second cut from turning into six.

Do not use it as a guarantee of a good clip. The README is honest about this: a perfectly formatted prompt still yields stochastic takes, and the fix is selective regeneration rather than rewriting everything. And do not carry a schema from one surface to another — the skill's whole point is that the API, the local diffusers/ComfyUI path and hosted providers each have their own rules, so a format that works through the API can be wrong locally.

Skip it too if you generate H3 through a provider that accepts free-form prompts. The references/freeform-en.txt format exists for exactly that case, with [0s-3s] blocks instead of the trained format.

What stays unverified

The README labels community findings about local LoRAs, speed tiers and long-video chaining as experimental in references/experimental-local.md, separate from the documented facts. It documents everything per interface with an evidence ledger of 18 primary sources at three confidence levels. Treat the experimental file as a starting point, not a specification.

The skill was built following the skill-creator structure and is MIT-licensed.

Sources

1. MiniMax H3 Prompt Writer — repository — modes, format fields, per-interface duration rules, structure table and honest limits quoted above.
2. README.md — install commands, usage examples, license.

Project page: MiniMax H3 Prompt Writer

For the format behind the scenes, see the MiniMax H3 multi-scene prompt guide, and for what the model costs per second at each resolution, H3 seconds per resolution.

FAQ

Which mode should I pick for a single still image? I2VA when the image is the first frame, or L2VA when the image is the last frame. T2VA is for text only, and Ref2VA is for loose style, motion or voice references rather than an exact frame.

Does the skill write the prompt in English? Yes. H3 is trained on structured English prompts, so the prompt itself is English even when the dialogue inside the <d> tags is in another language.

Why does the local frame grid say 17*n+5? Because that is the local frame-count rule the README documents at 24fps — it is not the same rule as the direct API's duration and ratio fields, which is why the skill keeps the interfaces apart.