Two months ago, a developer reverse-engineered the design logic behind Anthropic’s Claude Design and open-sourced it as Huashu-Design. Within weeks, the project amassed over 16,000 stars on GitHub. It could generate website mockups, app prototypes, presentation decks, even animated demos—all from a single sentence. But during a tutorial video recording, the creator noticed a troubling pattern: every output, while individually pleasing, converged to the same safe minimalism—white backgrounds, generous whitespace, a single accent color. The project had fallen into the trap that haunts most AI design tools: homogenous output disguised as quality.
The problem wasn’t unique to Huashu-Design. From Adobe Firefly to Canva’s AI, most generative design tools suffer from what researchers call "distribution capture"—models learn the most statistically common aesthetic and produce variations of it. Even with 20 built-in style options in v1, the agent kept retreating to the same Anthropic-inspired look. Faced with vague requirements, AI automatically retreats to its safest corner. The human designer might have made bold choices; the AI made safe ones.
To break this cycle, the creator rewrote the entire skill system from scratch, launching Huashu-Design 2.0 with three independent creative engines, each designed to actively fight monotony. These engines are best remembered by three verbs: Collision, Borrowing, and Invitation.
Collision forces randomness. The agent picks a style at random from an expanded palette, not to guarantee quality but to break habitual patterns. Sometimes a jarring choice leads to unexpected beauty. Borrowing grounds the output in real-world excellence. The agent searches for award-winning websites in the same category (e.g., Clio, Awwwards, CSS Design Awards) and migrates their design language—typography, color schemes, layout grid—to the new project. It doesn’t rely on imagination; it anchors on the highest standards of the real world. Invitation asks a provocative question: "If budget were unlimited, which legendary designer would you hire?" The agent then embodies that designer’s philosophy—be it Massimo Vignelli’s grid obsession, Paula Scher’s typographic energy, or Yayoi Kusama’s infinitive patterns—and produces a version from scratch.
The three engines ran in parallel for every request. When asked to create a promotional website for Malaysian tourism, the system returned three wildly different homepage drafts with just one prompt. A dark editorial style with explosive tropical images on warm black backgrounds, inspired by Monocle magazine. An immersive full-bleed photo layout reminiscent of the Clio Bronze-winning Visit Faroe Islands site. A warm-toned magazine style like Condé Nast Traveler. One dark, one bright, one serene—deliberately varied in temperature and mood. The most surprising was the editorial dark version, which the creator admitted they would never have chosen themselves. The true value of an AI design tool isn’t speed—it’s the unexpected directions it can suggest.
Another test produced a presentation on Japanese maid café culture, complete with pastel tones, decorative borders, and playful typography that matched the subculture’s aesthetics—far removed from any tech startup landing page. A third output designed a Chinese-language guided tour app for the British Museum, using a restrained museum-inspired palette of marble white, antique gold, and deep navy, with iconography referencing classical artifacts. Three projects, three entirely different visual languages, all generated in a single session with minimal human input.
The implications extend beyond one project. As AI-generated design becomes ubiquitous, the risk of visual homogeneity grows. Entire industries could converge on a handful of safe aesthetics, killing the diversity that makes graphic design vibrant. Huashu-Design 2.0 offers a replicable framework: explicitly introduce randomness, anchor to real-world benchmarks, and simulate human creative abstraction. It doesn’t replace the designer’s eye—it augments it by providing alternatives the designer might never consider.
For developers and designers experimenting with AI tools, the lesson is clear: the bottleneck isn’t generating something good—it’s generating something distinct. Good design shouldn’t have only one way of being. By adding friction and forcing variety at the generation stage, we can preserve the human joy of discovery inside the machine output. Huashu-Design 2.0 is open source and available on GitHub; its approach is already inspiring other projects like Open Design to rethink their diversity logic. The future of AI-assisted creativity may not be about doing things faster, but about doing things differently.