The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI
Treat an agent’s first UI as a draft: encode recurring design dislikes, supply visual references, split work into focused prompts, and reserve time for iteration.
Hallmark encodes recurring AI-generated UI patterns to avoid and supplies theme references; the speaker says **over 10,000 people** tried it. He recommends spending another **10–20%** on UI refinement and showed GLM 5.2 iteration beside an Opus result that cost **five times as much**.
Keep design preferences in a reusable skill or agent file, attach screenshots and references, and describe the user and interaction in detail. Build a base, then iterate feature by feature, using a sufficiently capable faster model where it fits.
Hallmark encodes recurring AI-generated UI patterns to avoid and supplies theme references; the speaker says **over 10,000 people** tried it. He recommends spending another **10–20%** on UI refinement and showed GLM 5.2 iteration beside an Opus result that cost **five times as much**. Keep design preferences in a reusable skill or agent file, attach screenshots and references, and describe the user and interaction in detail. Build a base, then iterate feature by feature, using a sufficiently capable faster model where it fits. The visual comparisons were selected demos rather than a systematic design evaluation. Avoiding familiar gradients or typography can remove obvious tells, but it does not by itself establish usability, accessibility, or a coherent product identity.
This confirms portable design skills as a practical way to steer agents away from recurring generic UI patterns, while adding an iterative workflow and a claimed cost-capability tradeoff between models. Against the prior candidates, it does not establish design quality as a universal checklist: selected demos and pattern avoidance still leave usability, accessibility, brand coherence, and subjective evaluation unresolved.