What's this about?
Together AI's Hassan shared his method for getting agents to design good interfaces: a `/Hallmark` design skill, at 25.8k stars, that enumerates every AI-slop design cliché as an explicit "don't do this" rule and feeds the list to the model. His other lessons: learn the AI clichés so you can steer the agent away from them; most importantly, feed the agent reference images and screenshots; make prompts longer and more specific; and iterate with a cheap small model — he reports GLM 5.2 being nearly indistinguishable from Opus for this, and much faster. Baoyu disagreed with the emphasis: the skill matters little. What matters is human taste first; then the model (GPT can't design, Claude is much better); and then a personalized design system with real conventions and components, rather than a pile of rules about what not to do.
What exactly is "AI slop" in design?
Generated interfaces have recognizable tells: purple-to-blue gradients, glassmorphism panels, the same three-card feature grid, oversized rounded corners, emoji as iconography, hero sections with a centred headline and two buttons. These appear because they're overrepresented in training data — they were fashionable, so they were everywhere, so the model treats them as what a website looks like. The result is competent and instantly identifiable as machine-made, which is why "avoid these" is a surprisingly effective instruction.
What's the community saying?
- The disagreement is the substance, and both are partly right. Prohibitions get you out of the default basin; a design system tells you where to go instead. A rule list can only ever prevent bad output, not produce a coherent visual identity.
- "Feed it reference images" was Hassan's own top tip, which quietly supports Baoyu — showing the model the target does more than describing what to avoid.
- The cheap-model finding is the practical takeaway: GLM 5.2 nearly indistinguishable from Opus for design iteration, at much higher speed, means the expensive model isn't the bottleneck for this task.
- "GPT can't design, Claude is much better" is a strong claim stated flatly by someone who does this daily, and it matches the split reported this week on vision and taste.