Key Takeaways

  • William Imoh built Mural, a real-time voice app for learning Norwegian, using Codex and GPT-Live without writing native Swift or Kotlin code.
  • Default code generation models lean heavily on generic dark mode layouts unless founders give explicit physical and aesthetic boundaries.
  • Imoh forced high visual quality by specifying real-world references: autumn color palettes, light mode, gradients, and liquid glass materials.
  • Aesthetic control requires strict negative constraints: Imoh instructed the model that no visual element could exist on screen without a direct functional requirement.
  • Claire Vo notes that design taste in AI development is not a mysterious gift, but the practical ability to use established design terminology and real-world references.

The Method

When founders ask an AI model to write software, the default output looks like an unstyled developer tool. It defaults to dark backgrounds, high-contrast borders, and clutter. Imoh avoided this trap while building Mural by treating the prompt like a creative brief for a human staff designer.

First, he set explicit role and atmospheric constraints. He told the model to act like a senior designer at Apple, but tied that persona to physical, real-world conditions rather than abstract praise:

“I literally wrote I want something that feels like fall, like autumn. Use fall colors. And again, it should be really beautiful. Follow design best practices like you were senior designer at Apple.” (play from 10:16)

Second, he banned dark mode directly. Default AI models tend to generate dark developer dashboards because code repositories are full of them. Imoh explicitly overrode that training bias:

“I was like, first this shouldn't be in dark mode. Not everything belongs in dark mode. It should be light mode. Use light warm colors. Explore gradients and use liquid glass as much as possible as well.” (play from 11:28)

Third, he applied extreme minimalism as a negative filter. Instead of asking the model to add elements, he commanded it to prune visual noise. The prompt established a baseline rule: “No element on the screen should exist unless it absolutely has to. And if something isn't necessary, she will remove it or put a toggle for it.” (play from 10:50)

Vo points out that this is where most technical builders fail with AI tools. Founders complain that AI lacks taste, but they fail to give the model the precise vocabulary needed to execute quality work:

“When people talk about taste they talk about taste like what is taste we need taste. And really what you need is design language and references in the world that you can point to. You need to know what liquid glass is. You need to know what minimalism is. You need to know what a gradient is. And I think the more rich your reference set is when it comes to design, the better.” (play from 12:21)

Where This Breaks Down

Atmospheric prompts work well for consumer interfaces that prioritize mood, tactile polish, and quick voice interactions. They struggle when software requires dense data presentation. Liquid glass, warm seasonal palettes, and aggressive element pruning break down on enterprise dashboards, billing reconciliation screens, and developer logs. In those products, hiding secondary actions behind toggles creates friction for power users who need high information density.

Furthermore, prompting an AI model to behave like an Apple designer works only if the founder knows enough design vocabulary to correct deviations. If a model generates blurry blurs instead of proper translucency, vague feedback fails. You must name the exact visual defect to fix it.

What to Do With This

Take the primary screen of your current prototype. Write a prompt that explicitly bans dark mode, specifies a light mode palette based on an analog material or season, and mandates liquid glass or translucent surfaces. Add a rule stating that any secondary element must be removed or moved behind a toggle. Run the prompt through your code generator and inspect the difference in native visual quality.