AI slop or AI cream of the crop?
Both come out of the same models. The setup around the prompt decides which one you get, and this site is the evidence.
"AI-made" tells you almost nothing about quality, because the phrase covers a prompt typed into an empty chat window and also covers an agent working inside a design system, a written rulebook, a memory of every correction I have ever given it and a test suite that fails the build when something slips. Those two produce work at opposite ends of the range, and people who have only seen the first one assume the second can't exist. Plenty of AI output deserves the word "slop," and I understand the reflex. What produces it is a missing setup, and a setup is something you can build.
What the slop end looks like
A one-shot prompt with no context, no memory and no checks returns something that looks finished in the preview. The buttons have the right radius for a generic button, the layout holds at the width the model imagined, and the colors are pleasant. Then you tab through it and the focus ring is missing, you open it on a phone and a row scrolls sideways, you switch to dark mode and a pastel card keeps its pastel under text that flipped to white, and you notice the dialog is a fresh hand-rolled one that ignores the three you already have. Nothing was checked, so nothing was caught, and the output looks done well before it is right. I wrote about why the defaults converge on the same look in defaults are the new template.
What the other end is built from
A 20+ person product team (designers, a design systems group, accessibility specialists, QA, front-end engineers, content writers) is mostly a machine for holding judgment in more than one head, so that no single person's tiredness or forgetfulness decides what ships. One person with AI can reproduce that output quality, because each of those roles has a concrete counterpart that can be written down and loaded into every session. This site is the evidence, so here is how each quality you would want from a serious product maps to a mechanism in the repo.
Hardened
The agent doesn't get to invent components. The project's CLAUDE.md orders its choices: first the 98 components in src/components/ui/, then the roughly seventy families of section templates in src/blocks/, then Radix primitives for anything interactive, then the 159 icons that all come through a single registry. Outside UI libraries are banned and so are hand-rolled dialogs, and a new component is only allowed when nothing fits, in which case it gets added to the system for every page to use. Every dialog on the site therefore has the focus handling and motion that were worked out once. The full anatomy of that file is in writing the CLAUDE.md for a design system.
Accessible
Color is where accessibility usually gets decided and then quietly undone, so every color token on this site comes from one source file and a generator writes the CSS from it. The test suite fails if any declared pairing of text on a background misses WCAG AA in either the light or the dark theme, and it fails if any CSS file gains a raw color literal. That second rule matters as much as the first, because it stops the agent from sidestepping the palette with a hex code it liked. Reduced motion is honored everywhere, and scroll-linked effects switch off completely under it. Editable fields are always 16px or larger so iOS Safari doesn't zoom the page when one takes focus. An accessibility specialist would catch these in review, and here a failing test catches them within seconds of the change.
Scalable
Scale means the hundredth page gets the same care as the first, so the care has to live in the system where nobody's attention can lapse. Every public route on the site is prerendered, and a motion effect that starts from opacity: 0 would blank a page that was already painted before any script runs, so after a build a grep for opacity: 0 in the generated HTML has to match nothing. A written rule in the design doc says how every entrance animation stays safe, and the grep confirms the rule was followed. The site is at version 1.43.0 with 126 published articles, and none of that growth required a rewrite.
Usable
A Playwright spec, e2e/responsive.spec.ts, checks every public route for horizontal overflow across phone, tablet, laptop and desktop projects, and it checks the navbar's menu and spacing at each of those widths too. Reviewing the result by eye is still part of the job, and I covered how I do it in evaluating AI-generated UI, but the review starts from a build that has already passed the mechanical checks, so my attention goes to the questions a person has to answer.
Beautiful
Beauty looks like the hardest one to automate, and it is the one that depends most on memory. The design doc (type scale, spacing, the elevation rules, the motion durations and springs) and a writing voice doc both load into every session. Beyond those, about 45 memory entries record my past decisions and corrections: a scroll effect I disliked and rejected stays rejected, and a gap in the design system gets raised with me before anything is substituted. Taste is a long list of small decisions made by a specific person, and once those decisions are written down the agent applies them the same way on day 200 as on day one. What makes an interface feel polished is mostly that consistency.
The setup is the work
None of this appears when you look at a single screen of output, which is why the argument over AI quality keeps going in circles. The skill is in what surrounds the prompt: the system the agent must reuse, the rules it reads first, the memory of what I already decided and the checks that fail loudly. Skills package the judgment, and the review work that remains is what the bottleneck moved to. Building that setup has taken deliberate work since March, and all of it was done by one person.
So, slop or cream of the crop? The setup around the prompt decides it. A design system the agent has to reuse, written rules it reads before every task, a memory of the decisions I've already made and checks that fail the build when something slips: together, those four let one person ship the polish of a 20+ person product team.
If you have only seen the slop end, try the other one on a single small project: write down three of your team's non-negotiables, make one of them a failing test, and watch how different the next output looks!
Be the first to rate this article.
Let's work together.
Open to select projects and collaborations — design systems, accessibility, and AI-native product work.