MiniMax H3 — Claude Code Skill

Current Reel Maker: two modes, ready-to-import workflows

For reference-based reels, install MiniMax H3 Reel Maker. It chooses the right branch from your inputs:

Your inputs	Workflow	What is required
An original video whose action should be reproduced	Ref + ControlNet + TS smoother	Source clip, face/body references and 1–3 scene anchors; aligned pose/depth controls
Photos and a description, with no original video	Ref Only	Relevant photos and text; no original video, pose/depth, ControlNet, smoother or source audio required

• Source-video workflow JSON
• Photos-only workflow JSON
• Setup, dependencies and commands
• Prompt structure and complete photos-only example
• Offline HTML quick start — download and open locally.

The JSON files are templates: replace the prompt and inputs before rendering, or let the skill's configure_workflow.py create a package with only the images you actually supply. Model weights and personal photo/video examples are not bundled. The source-video graph comes from a previously accepted render; the pure photos-only graph is locally checked but has not been GPU-rendered for this release. See the included provenance record.

Install the dedicated skill

Clone/download this repository, then copy the entire skills/minimax-h3-reel-maker folder into your agent's skill directory. Do not copy only SKILL.md: its scripts, references and assets are required.

• Codex: $CODEX_HOME/skills/minimax-h3-reel-maker, normally ~/.codex/skills/minimax-h3-reel-maker.
• Claude Code: ~/.claude/skills/minimax-h3-reel-maker, or the project's .claude/skills/minimax-h3-reel-maker.

Open a new turn/session so the agent discovers the installed skill. Ask, for example:

Use $minimax-h3-reel-maker. I have these photos and no source video. Create a short scene where the person waves naturally. Prepare Ref Only; do not request ControlNet or pose inputs.

Use $minimax-h3-reel-maker. Rebuild this source reel using my face/body photos and matching scene references. Keep pose/depth ControlNet and TS smoothing. Preserve the clothing, action and dialogue.

Existing session authorization still governs installs, GPU jobs and uploads. The preparation helper is local-only and never starts a render. Save every actual generation attempt separately.

General H3 prompting skill and advanced workflows

A Claude Code skill for MiniMax H3, the open-weight omni-modal video model that generates video and native stereo audio in a single pass. Install it once and Claude stops guessing: it picks the right checkpoint, writes prompts in MiniMax's documented rewrite-output structure, and knows why your camera move rendered a physical camera.

Not using Claude Code? The same thing ships as portable-prompt.md — one block of text you paste into ChatGPT, Grok or anything else. Building a whole reel rather than a single shot? reels-portable-prompt.md is the same knowledge wrapped in a clip-by-clip pipeline.

Built on MiniMax's own prompt-writing guides, plus the failure modes those guides don't cover.

---

Why