MiniMax H3 Prompting Skill
English | 简体中文
A guided, general-purpose Agent Skill for MiniMax H3 / Hailuo H3. It evaluates requirement maturity, H3 fit, and evidence strength; assigns information-density and multimodal-reference roles; then compiles the result into H3-compatible prompt fields with a built-in preflight.

What it solves
- Supports five modes:
T2VA,I2VA,FL2VA,L2VA, andRef2VA. - Uses local file reads only.
agents/openai.yamlsupplies optional ChatGPT / Codex UI metadata and does not restrict use by other Agents. - Selects task recipes for typography PV, Motion Design / MG, products, e-commerce, UI, game characters, character introductions, MV, title sequences, hand-drawn reality rewrites, audio reference, fashion lookbooks, and more.
- Budgets semantic states, editorial microcuts, graphic layouts, material effects, and sound attacks separately, so "information overload" is not mistakenly reduced to low density.
- Assigns each image, video, and audio reference one explicit responsibility, reducing competition between references.
- Separates official capabilities, inspected local outputs, and external examples instead of turning a single result into a universal rule.
- Does not import behavior from Jimeng SD2, Seedance, or other models into H3.
Workflow
Requirement maturity
→ H3 fit / evidence level
→ Information-density budget
→ Reference responsibility
→ Task recipe
→ H3-compatible field compilation
→ Preflight
Installation
Using the skills CLI:
npx skills add https://github.com/q2522879285-source/minimax-h3-prompting-skill-public --skill minimax-h3-prompting
Alternatively, copy skills/minimax-h3-prompting into your local Agent Skills directory.
Structure
skills/minimax-h3-prompting/
├── SKILL.md
├── LICENSE
├── agents/openai.yaml
└── references/
├── h3-interoperability-guide.md
├── guided-task-recipes.md
└── validated-findings.md
v2.7 update
- Adds a bounded attribute card × action proof character-introduction recipe.
- Uses a single-character chain: identity ignition, status label, action proof, impact climax, and hero lockup.
- Alternates bright MG definition fields with dark action fields to raise editorial density while keeping semantic load low.
- Treats readable-text whitelists, character continuity, action complexity, and native-generation versus post-production contribution as separate review surfaces.
- Uses external examples as structural evidence only—not as proof of complex combat, exact long-form text, or native H3 compositing.
Evidence boundary
This project uses public MiniMax pages to check interface facts without redistributing their documentation text. It distinguishes:
- officially supported capabilities;
- inspected local outputs;
- external examples or unvalidated design candidates.
An external prompt, post, or one-off result may become a bounded design recipe, but never an automatic claim about H3's general capabilities.
Upstream
This is an independently maintained community project and is not affiliated with MiniMax. MiniMax, Hailuo, and related names belong to their respective owners.
License
MIT. This repository does not redistribute MiniMax documentation text. H3 mode names, interface keys, and machine identifiers are used only to describe compatibility.
No comments yet
Be the first to share your take.