MiniMax H3 Official Skills Getting Started Guide: 9 Ready-to-Use Video Generation Skills

Introduction: The Hardest Part of Generating Videos Isn’t Pushing the Button — It’s “Translating”

Anyone who has used AI video tools will likely feel the same way: what really stalls you is never the act of “clicking generate,” but rather how to translate the image in your head — lighting, camera movement, pacing, beat-syncing, sound — into a language the model can understand.

Between an average user and a product ad with “Apple keynote quality” lies hundreds of words of storyboarding, camera language, transition design, and music sync. In the past, this either required endless trial and error, or relied on the intuition of seasoned hands. MiniMax recently did something very clever: it took the question of “how to write high-quality prompts” and packaged it directly into official Skills that can be installed and reused.

This article won’t dive into obscure technical architecture. It covers only one thing: MiniMax H3 officially provides you with 9 skills — what each one does, how to install them, how to use them, and which pitfalls the official team has already flagged for you in advance.

Core takeaway in one sentence: H3’s value lies not just in the model itself, but in the fact that MiniMax has “productized” professional video production workflows into 9 installable Skills — install to get an official workflow, and even ordinary people can reproduce professional-grade output like Apple-style ads, 3D animations, and stop-motion explainer videos.

1. Getting to Know H3: What It Is in One Paragraph

MiniMax H3 is an open-source omni-modal generation system from MiniMax. It can simultaneously understand mixed inputs of text, images, video, and audio, and directly generate videos with native stereo sound built in.

Key specs from the official team:

• Output duration: 4–15 seconds

• Supports multiple aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, etc.

• Default short edge 768p, up to 2K via H3-Regenerate-2K

• 24 FPS, 32kHz stereo audio

• Stable support for 11 languages of conversation (including Chinese and English), with varying degrees of support for other languages

For the average user, there are really only two things worth remembering: the videos it generates come with sound built in (no need to hunt for BGM and sound effects separately), and it supports multimodal reference input — images, videos, and audio can all be fed in as source material.

And the skills directory released alongside the repo is precisely what enables ordinary people to make good use of this capability.

2. The Official Skills Lineup: A 1 + 8 Structure

Open the skills directory in the repo, and you’ll find 9 skills, structured as “1 core + 8 styles”:

1 core prompting skill:

• h3-prompt-writing — rewrites any multimodal need into H3’s standard prompt structure. It’s the foundation for all advanced use cases.

8 style-based video generation skills (all with bilingual documentation in SKILL.md / SKILL.cn.md):

Note one detail: all 8 style skills are bilingual in Chinese and English, but the core h3-prompt-writing is currently only available in English per the official team’s note — it serves the task of “writing prompts,” and the skill is still being iterated. In practice, however, it handles Chinese-language needs perfectly well, and can even fill in incomplete prompt templates from the official docs very effectively (see the “Hands-On Experience” section below).

3. The Core Skill h3-prompt-writing: Five Modes and Three-Part Prompts

This is the most “hardcore” of the 9 skills, and it deserves its own section.

H3 supports five video generation modes. The role of h3-prompt-writing is to help you determine which mode your current need falls into, and rewrite the prompt according to the corresponding structure:

• T2VA: Text only → build a complete audiovisual timeline (text-to-video)

• I2VA: Start from a first-frame image and extrapolate forward (first-frame-to-video)

• FL2VA: Describe the continuous path between a first frame and a last frame (first-and-last-frame-to-video)

• L2VA: Infer a plausible opening that converges to the given last frame (last-frame-to-video)

• Ref2VA: Full-reference mode, supporting mixed image, video, and audio input (multimodal reference-to-video)

For the base modes (T2VA/I2VA/FL2VA/L2VA), you need to rewrite three core fields in the official order:

• integrated_multimodal_description — the integrated multimodal visual description, detailing composition, subject, environment, action, and camera movement shot by shot

• overall_soundscape — the overall soundscape, covering the sound corresponding to each action in the frame

• non_diegetic_music — non-diegetic music, i.e., the tonal direction of the background music

The full-reference mode Ref2VA is more complex, requiring rewriting across six parts: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, and non_diegetic_music.

The real value of this skill: the official team has codified “what it takes for H3 to understand you” into executable rules — for example, requiring each shot to be described in terms of “composition, subject, environment, action, camera movement, sound, and the exact location where reference content appears,” requiring timestamp annotations to match the requested duration, and requiring dialogues and lyrics to be preserved in their original language. In other words, it’s a set of “officially certified prompt best practices.”

4. The Eight Style Skills: Breaking “Directing” Down into Executable Steps

What these style skills share is that they don’t shoot the video for you — they break the director’s decisions down into explicit steps for you. Take minimalist-product-ad-generator as an example. Its full workflow is:

• Opening questionnaire: first confirms product material, style, target duration (10 seconds is the default recommendation), aspect ratio, Apple-style template, and copy approach — this is a mandatory gate

• Product fact summary: analyzes category, dominant colors, and displayable structure

• Choose the narrative spine: product reveal / tactile feature / color family

• Define motion language: specifies movement intensity, transition logic, and rhythm peaks

• Generate English copy: 3–5 word Apple style, no template recycling

• Three independent anchor photos: lock in the hero angle, material detail, and closing copy composition

• Precise beat-based storyboard table: an execution sheet for the video model, planning shot, action, and text second by second

• Video generation: MiniMax-H3 by default, with native audio

• Music and beat-synced editing: re-score with music-2.6 if the music isn’t a good fit

• Delivery verification: checks aspect ratio, copy, product fidelity, and music item by item

Note that many of these details are lessons baked in after the official team hit pitfalls themselves, such as: don’t use four-quadrant anchor grids (video models carry the grid layout into the final output), on-screen text needs integrated motion design rather than relying on post-production subtitles, and the product’s actual color is a hard fidelity constraint (Apple-style doesn’t mean dyeing the product white)… These rules read like a “product director’s manual.”

The other style skills follow the same philosophy: 3D animation goes from project brief all the way to storyboard, per-shot generation, assembly, and BGM; the stop-motion explainer first establishes learning objectives and visual metaphors, then moves to paper puppets, dimensional sets, and storyboarding; the hand-drawn + live-action skill even checks for “whether the contact is physically real, whether camera movement feels laggy, and whether the tone is non-horror.”

5. How to Use: One Command to Install, Four Entry Points to Start

Installing skills uses Vercel’s skills CLI, with a single command:

npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing

Swap in another skill name (e.g., minimalist-product-ad-generator) to install the corresponding style skill.

There are four entry points for using H3 itself:

• Online API: Global platform.minimax.io / China platform.minimaxi.com

• Online App: Global hailuoai.video / China hailuoai.com

• Desktop: Global hub.minimax.io / China hub.minimaxi.com

• Local deployment: H3-Base weights are open-sourced (Hugging Face / ModelScope), with support for SGLang, vLLM, diffusers, ComfyUI and other frameworks (hands-on reminder: local deployment involves high hardware costs and a steep configuration bar; cloud deployment is the better recommendation for saving money and hassle — see Section 8)

For the vast majority of content creators, the most efficient path is: install the skill → follow the steps the skill gives you in the App or API. The skill standardizes “how to formulate your request”; all you need to do is provide the materials and make choices.

6. How Different People Should Choose

• E-commerce sellers / brand owners: go straight to minimalist-product-ad-generator — start with a single product image, Apple-style by default

• Creators who want animated shorts: 3d-animation-short-generator gives you a complete pipeline from concept to finished film

• Science / education bloggers: papercraft-stop-motion-explainer or paper-collage-explainer-generator — two handcrafted visual styles to choose from

• Musicians / MV creators: music-video-subtitle-generator — lyric typography and beat-syncing are its strengths

• Game concept / social media creators: co-op-game-intro-generator can create two-player game intro animations

• People who want to write high-quality custom prompts: master h3-prompt-writing first

7. Boundaries and Truths: The Pitfalls the Official Team Has Already Flagged

It’s worth noting that every official skill honestly declares what it’s “not suitable for”:

• Minimalist product ads: not suitable for KOC spoken reviews, plain editing, or complex screen demonstrations

• 3D animation: not suitable for single still images, simple cuts, photorealistic humans, or single clips

• Hand-drawn + live action: not suitable for fine CG, horror/scare content, furry characters, or multi-scene editing

This boundary information matters a lot — it means the official team doesn’t promise “one skill handles all needs,” but rather wants you to choose based on the scenario. Also, the skills are still under active iteration, and h3-prompt-writing is currently English-only for now. If you optimize an existing skill or add a new one, you can submit a PR — the official team explicitly states that contributing and improving skills earns API rewards.

One final reminder: when generating videos with H3, input content goes through safety review, and material involving illegal activity, pornography, or infringement may be blocked — set your expectations accordingly.

8. Hands-On Experience: Is It Actually Good?

This section is based on real hands-on impressions. Straight to the conclusions.

Prompt completion: what the official docs didn’t fill in, it completes very well. Testing h3-prompt-writing (including Chinese-language use cases), the most pleasant surprise is this: even when some prompt examples in the official documentation are incomplete, the skill still follows H3’s five-mode structure to reasonably fill in the missing parts, and the resulting video style is very much on point. In other words, you don’t need to become a prompt expert yourself — just hand it your requirements and it produces prompts that meet the spec and can be fed directly into H3.

Deployment cost: don’t build your own hardware — going cloud saves you more. If you’ve had thoughts about locally deploying H3-Base, the hands-on conclusion is: hardware costs are high, and it’s not friendly to individuals or small teams. Cloud deployment is the stronger recommendation, with much better cost-effectiveness — generating a 15-second video costs roughly 0.6 RMB. We recommend prioritizing the official accelerated deployment version, which offers better speed and cost.

Closing: Get Started

From today on, you no longer have to write prompts from scratch. MiniMax H3’s 9 official skills are, at their core, a set of “officially certified video production SOPs” — install one, follow its steps, and all that’s left is handing your material and ideas to the workflow.

Your first action: pick the skill closest to your business (e.g., if you’re in e-commerce, choose the minimalist product ad), and run the full pipeline once. You’ll be surprised to find that a finished piece that “looks like it needs a professional team” is really just one command and a few confirmations away.

This article is compiled from the skills directory, README, and per-skill SKILL.md files in the official MiniMax H3 repository (github.com/MiniMax-AI/MiniMax-H3). All skill names, modes, and steps in this article are from the official documentation.