Source: raw/ChatGPT_+_MiniMax_Design_-_How_to_Make_Motion_Graphics_with_AI_Full_AI_Workflow.md (https://www.youtube.com/watch?v=TrSjEBArbD0), YouTube auto-captions fetched 2026-10-30. Creator: AI Master (youtube.com/@iamAImaster) Platform: YouTube Published: not recorded in the raw file

Sponsored product demo

“This video is sponsored by MiniMax Design.” Every step runs inside the sponsor’s product, and the cost figure below is MiniMax’s own. The reusable part is the order of work (references, then a locked style brief, then a shot plan, then generation), which applies in any multi-model tool.

MiniMax Design (design.minimax.io) is an AI creative studio: one canvas holding image, video, audio and text generation, with an agent on top that plans the job and calls the right model for each step. The creator builds a motion-graphics package for a product explainer (a 10-second kinetic-typography opener and two short bumpers) with GPT-6 Astra planning in agent mode, GPT Image 2.5 drawing style frames and MiniMax H3 generating the video, then edits the result in the tool’s built-in editor. His verdict: it “earns a spot in the front half of a motion pipeline” and does not replace After Effects or Blender.

Key Takeaways

  • One workspace, several models. The agent “works with your local files, reads web pages, and calls the right model for each step.” In this build: GPT-6 Astra for the plan, GPT Image 2.5 for style frames, MiniMax H3 for video, plus sound and editing in the same canvas.
  • The style brief is the step that saves credits. Before anything is generated, the agent studies the reference board and writes a short brief as editable text on the canvas. He fixes one paragraph, matches the accent colour to the brand kit, and locks it. Every later model call pulls from that one source of truth.
  • Agent mode plans like an editor, not a chatbot. The shot list came back with timings, frame counts per shot, copy short enough to read in a second, and a reason for each shot, detailed enough to build the After Effects comp from. Revisions rewrite only the shots you name.
  • On-screen copy starts from the product’s own claims. Paste the landing page; the agent pulls claims “six words or less each,” so the model “doesn’t get to invent features.” He still checks every line against the client’s approved messaging.
  • H3 held moving type. “Launch stays spelled correctly through every slam and settle,” edges sharp mid-overshoot, and all three parts of his animation note (one-by-one entrance, slight overshoot, one-direction stripe wipe) appeared. He would still retime the easing in After Effects.
  • Two ways to animate an approved frame. Attach it as a style reference, or feed it as the start frame (image-to-video) when the client wants the composition untouched.
  • 15 seconds per clip. H3 generates up to 15 s per clip. Plan longer pieces as a series of short shots, review each, and stitch them.
  • Cost claim (vendor). “MiniMax says running GPT-6 Astra through MiniMax Design is about 92% cheaper than running it on GPT directly. That’s their number, so check it against your own usage.”

The Workflow, Step by Step

  1. Reference board. Use the browser built into the workspace; search (here: bold typographic poster layouts, film-grain textures) and drag finds straight onto the canvas. He used only free-licensed texture libraries and the product’s brand kit: “A reference you don’t have the rights to is a reference you can’t ship.”
  2. Pull copy from the live page. Paste the product landing page into the chat and ask for on-screen claims of six words or fewer.
  3. Style brief. Ask the agent to study the references and write a short brief before generating anything. Edit it, match the brand accent colour, lock it.
  4. Plan in agent mode. Switch to GPT-6 Astra, turn on agent mode, and brief the whole deliverable in one prompt “the way I’d brief a junior motion designer.” Check that it returns timings and frame counts, not mood words.
  5. Revise the plan in chat. He asked for punchier copy on bumper two and a faster cut mid-opener; only those shots changed.
  6. Style frames with GPT Image 2.5. Type-heavy frames are his stress test (“big letters expose every weakness”: spelling, kerning, baseline). The frames spelled the product name correctly and placed the accent stripe where the brief said. He kept two of four and sent two back with one note each; regenerations stayed in the look because the brief lives on the canvas.
  7. Video with H3. Write the motion as an animation note: “The letters slam in from the bottom one at a time, overshoot slightly, and settle, and then the accent stripe wipes left to right… over about 4 seconds.” Attach the approved frame as a style reference.
  8. Check output dimensions before anything goes into a comp: “the aspect ratio doesn’t always match” what the prompt asked for.
  9. Sound. Ask the agent for a whoosh-and-hit pass timed to the letter slams and a low synth bed for the bumpers. It reads its own shot list for timing. A couple of hits landed “early by a frame or two”, a 30-second fix.
  10. Edit in the built-in editor. Clips go onto a timeline in the workspace; it adds subtitles automatically. For frame-by-frame finishing, export to After Effects.

Other Features Shown

  • 3D Director Stage (add node on any canvas): a simple 3D scene where you pose character bases and place and schedule cameras. It replaces the rough previs pass he would block out in Blender for a client, not final renders; its frames feed back into generation as references.
  • Blender connection: with it linked, GPT-6 Astra in agent mode can set up or change your Blender scene from the same chat.
  • ComfyUI integration: a local ComfyUI deployment and its workflows can run inside the canvas. MiniMax also says workflows from other platforms can be ported in and that H3 can run locally. He had not tested either enough for a verdict.
  • Story pieces: for a film-style intro or short drama, ask Astra in agent mode for a script, scene breakdown and shot list first; MiniMax says story and script planning is one of the things the agent workflow was built for.
  • Commercial use: MiniMax positions H3 for professional motion graphics. “Check the commercial use terms on your plan before you deliver it.”

Try It

  1. Run it on the job you do most often, not a demo prompt: “a real brief, a real brand kit, and a real deadline” (his advice).
  2. Make the style brief your first generated artefact in any multi-model tool, and lock it before spending credits on images or video. ^[inferred: the step is tool-agnostic; the same pattern works in Higgsfield or Claude-driven pipelines]
  3. Use a type-heavy frame (product name set huge) as your stress test for any image or video model.
  4. Write video prompts as animation notes: entrance order, overshoot, settle, direction, duration.
  5. Keep each shot at 4 seconds or less and plan longer pieces as separate shots under the 15-second cap.
  6. Before a paid client delivery, read the plan’s commercial-use terms and verify the 92% Astra saving against your own bill.

Implementation

Tool/Service: MiniMax Design (design.minimax.io), with GPT-6 Astra, GPT Image 2.5 and MiniMax H3 available inside it. Setup: Sign up at design.minimax.io; start from a blank canvas; use the built-in browser for references and the chat’s agent mode for planning. Cost: Not stated in the video beyond MiniMax’s claim that Astra through MiniMax Design is about 92% cheaper than direct. Agent runs “can fire off a lot of model calls in the background.” Integration notes:

  • H3: up to 15 s per clip; text-to-video with a style reference, or image-to-video from a start frame.
  • Connects to Blender (agent edits the scene) and ComfyUI (local workflows run in the canvas).
  • Export clips to After Effects or another editor for finishing.

Open Questions

  • MiniMax Design’s plan prices, credit costs per model call, and the basis of the 92% Astra figure are not given.
  • Whether the H3 inside MiniMax Design is the same model as the open-weights H3 covered in the August video-model wave, and how H3’s licence terms (flagged there as “kind of a strange license”) apply to client work.
  • The ComfyUI porting and local-H3 claims were not tested in the video.
  • The full planning prompt and the style-brief prompt were shown on screen, not read aloud, so they are not in the transcript.