Anshu Chimala — 12 years leading software engineering and design at Apple, now posting AI-design demos on X — argues the reason AI churns out “generic slop” for you and “magic” for him isn’t the model. LLMs are next-token predictors trained to make the safe, average choice every time, which is design-by-committee. The fix is a reimagined Double Diamond process — Discover → Define → Deliver — tuned for a team of AI agents.

Why AI design is bland

  • LLMs predict the most likely next token; feedback training rewards the choice that pleases everyone, so design collapses to the average
  • “Great design starts with feeling” — it bends the rules and aims for an emotional response, the exact opposite of an LLM’s natural behavior

Discover — go broad before deep

  • Seed strings: a model can’t act randomly, so inject variety from outside — have it generate a random string and treat it as design inspiration (String Seed of Thought, from Sakana AI)
  • Be more ambitious: bring your own taste — pixel-art, isometric 3D city, “break all the rules” prompts — then use AI to brainstorm broad, refine your favorites, and write the final build prompt
  • Save the prompts that fail and retest them when newer models ship

Define — give each design its own identity

  • Design-critic subagents: have the coding agent ask a fresh-context critic to score a screenshot out of 10, and stop only at 9/10 — a big model for taste, a cheap one for the grunt work
  • Image generation: agents avoid images (gradients and shapes are the tell) — wire in image tools, or hand the agent an API key with a tight spend limit
  • Video generation: loop a clip and matte out the background, or interpolate keyframes for scroll-scrubbed transitions (via an aggregator like fal.ai)

Deliver — polish by removing

  • AI loves to add and rarely takes away; restraint reads as premium
  • Cut the glows, gradients, redundant labels, and custom widgets that look worse than native iOS
  • “Putting less on the screen communicates more”

The through-line: taste is a steering problem, not a model problem. The same agent that emits cookie-cutter layouts when asked to “make something unique” produces one-of-a-kind work once you supply the randomness, the opinion, and the critic it can’t generate for itself.

“Most people only see 1% of AI’s creative potential. I want to show you how to tap into the other 99%.” — Anshu Chimala