Rexan Wong's Motion Graphics Prompt Workflow (Claude Opus 5.5)
Rexan Wong's Motion Graphics Prompt Workflow (Claude Opus 5.5)
Everyone's sharing motion graphic videos Opus 5.5 made, and it's genuinely insane. Everyone says they made it with "one prompt" — but a bare one-prompt attempt looks mid. Here's the workflow that actually gets pro-level results. — @rexan_wong
Links
- Tweet: https://x.com/rexan_wong/status/2103707054108299437
- Author: @rexan_wong
The Workflow
-
Reference a style, don't describe one. Pull 1–2 reference videos from whatships.com and tell Opus to match them by name. Naming a style beats describing it — without a reference, Opus falls back to its default look (centered text, gradient background, everything fading in), which is why so many AI motion graphics videos look the same.
-
Give Opus a renderer, not just a canvas. Install HyperFrames or Remotion so Opus writes each scene as code and renders straight to mp4 — no video editor, no screen-recording an HTML page. Every frame is exact, and a requested change edits one line and re-renders instead of starting over.
-
Use real UI components, not invented ones. Install 21st.dev for real buttons, cards, and components built by design engineers. Without it, Opus draws your product's UI from scratch — wrong spacing, placeholder boxes, fake-looking buttons — and the whole video reads as cheap.
-
Dump full context, then ask for 3 storyboards. Brand assets (logo, colors, fonts), real product screenshots, the chosen reference video, and a braindump of the vision — then ask for three storyboard variants. Without this, Opus guesses your colors, your fonts, and what your product even does. Three variants means picking a direction instead of patching the first idea it had.
-
Approve a still frame per scene before anything animates. Fixing a storyboard is far cheaper than fixing a render. Catching a wrong scene 4 only after the whole thing is animated means a full re-render; a still frame takes seconds to change.
-
Direct with camera language, not vague notes. "Slow every zoom to 0.7x," "hard cut here," "push in on the button" — specific, director-style notes get the last 20% that makes a render look pro. "Make it better" gets random changes.
Why Interesting
Everyone has access to the same model — the context and tooling wrapped around it (a reference video, a code-based renderer, real components, a staged storyboard review) is what actually separates "one-prompt mid" from pro-level output. A reusable pattern for agent-driven creative work generally, not just video.