How to keep AI-generated interfaces from looking the same
Do not ask AI for one version. State the real task and the constraints that cannot change, have it build directions that differ in their tradeoffs, then compare them and decide yourself. Below is how we do it, using the Pulse case. It comes from recorded runs, but we have not run a controlled user study, so it does not prove the result is better.
Published 2026-11-021. Name the reader and the task
“Build a dashboard” gives AI nothing to decide with. The Pulse brief reads: redesign the weekly analytics report of a subscription app for the product lead on Monday morning.
2. State what cannot change
Pulse marks its brand as locked in src/tokens.css. In the 2026-10-27 runs, both Claude Code and Codex recognized this, and the variant files declared no other fonts. The differences then land in structure, not in a new color scheme.
3. Make each direction optimize something different
Three versions that only change visual style are often one idea three times. Pulse uses the same week of data and the same brand for three ways of reading:
- Lead with what to act on
- Compare metrics side by side
- Trace why something changed
4. Borrow structure, not appearance
When you give a reference, say what to study and what to avoid. Study its information hierarchy and tradeoffs; do not copy its brand, text or assets.
5. Ask for a recommendation, then decide yourself
Have AI state what each direction optimizes, what it gives up and where it fits, then recommend one with reasons. A recommendation is not a choice; that belongs to someone who knows the product.