how did /design fix the ai design slop problem

the /design skill in Command Code can fix ai design slop https://CommandCode.ai ok so this one is a fun deep dive. the topic is "AI design slop," that very specific look you instantly recognize once someone points it out: the indigo gradient (it is indigo, not purple), the centered hero, three identical feature tiles, the unearned glassmorphism, the little icon-topper above every heading. once you see it you cannot unsee it. the punchline that I find genuinely interesting: this is not a capability problem. the model already knows OKLCH, knows reduced-motion, knows the ~65ch line-length trick, knows golden-ratio spacing. it can write CSS fluently. what it lacks is the policy, knowing when to deploy what it already knows. taste, it turns out, is not represented well in the statistical prior of the internet. so when you ask for a landing page the model samples the mode of all landing pages, and the mode is mediocre by definition. the part that surprised me most is how small the failure distribution is. you talk to good designers, ask them to label AI-generated UIs, and the tells collapse to roughly 10 things that account for ~90% of the "this is AI" signal. narrow failure distribution means a tractable, almost deterministic repair, the same shape as the tool-call repair work I did earlier on open models. a few ideas I keep coming back to in here: the real bug is compositional, not cosmetic. the model chooses layout before it chooses purpose. a dashboard is a monitor surface, a landing page is a decide surface, and they want completely different compositions, but the model reaches for centered-hero-plus-cards for both. forcing it to commit to a structural frame (which of 7 surfaces is this?) before generating any visual tokens is just chain of thought for design. same reason CoT helps with math: you reduce the entropy of generation by conditioning on a high-level plan first. OKLCH over HSL is a representation lesson. HSL lightness is perceptually nonlinear, so equal numeric steps give unequal visual steps and the model optimizes over a bumpy landscape full of discontinuities. OKLCH is perceptually uniform, so distance in parameter space tracks distance in output space. the general principle: when you design an interface for an LLM to operate, whether color space or schema or API, pick the representation where the gradient actually points somewhere useful. state coverage is the most honest metric we found, zero subjective judgment required. humans ship ~7 to 9 states per component (idle, hover, focus, active, loading, empty, error, disabled, overflow). the median AI component ships ~1.5. you can just count. laziness is measurable. and the meta point: what got built here is basically a reward model for design, implemented in structured English rather than weights. a rubric, a set of negative-reward "smells," a completeness check, a surface prior. ~4500 lines, and that is apparently enough to pass designer review. which suggests UI taste is lower-dimensional than it feels. not an infinite space, a finite catalog of principles plus the logic of when to apply them. the model did not get better at design. the harness did the teaching. same lesson as tool calling: not a capability gap, a contract gap. CHAPTERS 00:00 the AI design slop problem 00:35 intro: Command Code and the hardness tool-repair backstory 01:48 the /design skill, bundled in Command Code 04:44 LLMs write fluent CSS, zero design taste 07:17 why the average landing page is just the mode of the distribution 10:39 the failure data set is tiny (~10 tells) 14:52 live /design smell audit on the word-counter site 18:44 the real bug: layout chosen before purpose 19:33 seven surfaces as chain of thought for design 23:08 validate then repair: separate diagnostic from treatment 28:43 OKLCH vs HSL and why representation matters 30:08 pick emotion before hue, no more default indigo 30:51 state coverage as the most honest metric 33:00 killing the infinite repair loop 35:43 truthful completion (the hardest constraint) 37:40 a reward model for design, in structured English 41:02 a design co-pilot for developers (docs + blog) 41:53 applied example: the DeepSeek deal ticket 43:45 what's next: /design d-slop, and outro --- Follow on 𝕏: https://x.com/MrAhmadAwais Command Code: https://commandcode.ai GitHub: https://github.com/AhmadAwais LinkedIn:   / mrahmadawais   YouTube: https://YouTube/AhmadAwais Dev Blog: https://AhmadAwais.com If you like my work, feel free to share it, like it, and subscribe to my YouTube channel. Let's connect on X @ https://x.com/MrAhmadAwais Use your code for good. Peace! ⌘