Overview
How might we design generative ui so people can trust and act on AI output?
When to use
- Perfect for rapid prototyping tools, design systems, dashboard builders, and low-code platforms where speed from concept to UI is critical.
When to skip
- Simple Q&A where text is enough and generated widgets add noise.
- Strict design-system products that cannot safely render arbitrary generated components.
- High-security forms where model-authored fields would be unpredictable or unsafe.
Rules
Generating non-interactive mock UI that looks clickable but does nothing.
Inconsistent components that break layout or accessibility every turn.
No way to fall back to text or copy structured results out of the widget.
Regenerating the whole UI on every keystroke with no stable state.
Evidence
| Product | Implementation |
|---|---|
| ChatGPT | Can render interactive canvases, tables, and app-like blocks inside chat. |
| Gemini | Produces structured maps, tables, and actionable layouts for research and planning. |
| Vercel / AI SDK demos | Streams generative UI components tied to tool results. |
| Cursor | Surfaces interactive diffs, plans, and structured agent artifacts beyond plain chat. |
