MiniMax Design pairs H3 with video-production agents

• MiniMax Design turns a creative brief into a production pipeline around H3, with the full-pipeline announcement saying its agents plan, execute and deliver the work.
• An image-led request can become a camera and pacing brief before it reaches H3, as CharaspowerAI's FPV test demonstrates.
• Reusable video prompts are arriving as production schemas, with techhalla's paper-cut prompt locking a 15-second story into four timed acts.
• Motion graphics are the early showcase category, which the motion-graphics announcement names as H3's purpose and bennash's “Nash & Bucky” sequence puts into a 60-second title treatment.

MiniMax shipped the desktop workbench on August 20, according to Cocoloop's launch report. Its official workflow separates copy, image, video and audio workers, while MiniMax's H3 prompt-writing skill can also run inside outside agent harnesses. The interesting production detail is not the one-line FPV request, but the much longer direction brief behind it.

MiniMax says its main agent reads a brief, breaks it into tasks and automatically chooses models, with manual model selection still available. The official Design workflow describes this execution sequence:

1. Understand the creative goal.
2. Decompose the task.
3. Run four workers in parallel:
4. Merge, edit and export the result.

The company's full-pipeline post reduces the starting input to a goal, while its H3 and agent post frames the product as H3 paired with agent execution.

Design's product site divides the surrounding workspace into five stages:

• Describe the idea: submit a short direction or a full brief.
• Build the canvas: keep script, storyboard, video, music and editing in one node-based workspace.
• Use Skills and plugins: create a custom Skill in chat or load a prebuilt workflow.
• Manage local assets: save canvas assets locally, let agents access local files, then export to professional tools.
• Review before delivery: direct the agent at checkpoints while its Harness runs multi-round quality checks.

The company has pitched that surface toward After Effects, VFX and post-production in its production post, and its anime-PV post points to music-video making as an early packaged use case.

CharaspowerAI started with one image and the instruction, “Act as an expert FPV director and turn this into something viral.” The creator said Design analyzed the idea, then produced an H3 prompt covering:

• camera logic,
• movement,
• pacing,
• and the detailed shot direction needed for the generated run.

The visible command was short, but CharaspowerAI's follow-up says the brief given to the system ran beyond 6,000 characters, with rules for producing FPV shots from any type of image. That makes the workflow closer to handing off a director's packet than tossing a casual prompt into a video model.

techhalla's template turns a 15-second paper-cut narrative into a fixed technical frame, then leaves the story beats replaceable. Its reusable constraints are:

• exactly four visual acts, each with a timed interval;
• a mostly locked wide shot with occasional lateral drift;
• hard stop-motion jumps between acts;
• tactile paper, cardboard, ink silhouettes, layered collage and visible texture;
• a negative prompt barring smooth motion, morphing, dissolves, photorealistic characters and digital effects.

The creator's alternate template applies the same frame to dinosaur extinction, including a four-beat storyboard and the same camera restrictions. MayorKingAI's Trojan War clip shows the paper-cut treatment holding together across several shifts in action and framing.

Hailuo_AI repeated its motion-graphics positioning in the H3 motion-graphics post. The creator work is already splitting that capability into distinct jobs.

• Opening titles: bennash used the desktop agent for workflow planning and execution on a 60-second spy-film title sequence, while bennash's workflow reply says the image and music prompts were still written manually.
• Dialogue shots: CharaspowerAI's dialogue test calls out voice delivery, timing, facial expressions and acting in a direct-to-camera character clip.
• Type-led VFX: CharaspowerAI's “IGNITE” VFX prompt specifies a fire vortex that forges a word before exploding it into a shockwave, a prompt built like a title-reveal shot description rather than a style tag.

H3's official model page accepts text, images, video and audio as a shared creative context, and generates 5-15 second clips with native stereo sound. MiniMax's H3 launch post sets the maximum output at 2K, so Design's timeline, editing and export layer carries the job of turning individual clips into a finished production.

The Design site lists a free starting tier and desktop downloads for macOS 13 or later, on Apple Silicon or older Intel Macs, plus Windows 10 or later on x64. MiniMax's GitHub repository also ships an h3-prompt-writing Markdown skill, with reference guides for text/keyframe and full-reference modes, that it says works in Claude Code, Cursor, Codex and other compatible agent harnesses.

bennash reported that H3 moderation rejected a reference image and variations of it. Hailuo's model page advertises reference control over characters, motion, cameras, voices and editing style, but this report puts an input-moderation gate ahead of that workflow.