You have seen the meme: a grid of screenshots, a dozen AI-built SaaS sites, indistinguishable. Dark hero, centered headline set in Inter, a violet glow behind the product shot, three rounded cards, a gradient on the button. The founders are different. The products are different. The interface is one interface. The instinct is to blame the prompt: surely a sharper brief would have produced a sharper page. It would not have, and understanding why is the actual design lesson.
- The sameness is arithmetic. A model trained on the web’s landing pages predicts the mode of that distribution, and the mode is a shadcn card under a violet glow.
- Anthropic’s name for the failure is distributive convergence: the model reproduces the statistical center of design decisions instead of committing to a point of view.
- Better prompts move outputs slightly within the same distribution every other builder is sampling. They cannot leave it.
- The working escape is a constraint layer the model generates inside: banned defaults, owned tokens, named type. Taste-Skill’s 75.8k stars prove the demand; an envelope of your own is what proves identity.
The mode has a look
A language model asked to design an interface does what it does with every other request: it emits the most probable continuation. Its training data is the public web, and the public web’s frontend corpus is dominated by a specific artifact, the venture-backed SaaS landing page, built to one playbook for a decade. When you ask for a dashboard with no further direction, you are asking for the expected value of that corpus. The expected value has Inter at 600 weight, an indigo-to-violet gradient, and a 12px border radius.
The tooling compounds it. As practitioner reviews have documented, shadcn/ui plus Tailwind became the default AI coding-agent stack for a structural reason: their predictable, utility-class, copy-paste component patterns are exactly what the models saw most and generate most fluently. v0 emits shadcn by default. Cursor scaffolds it. Claude-family agents reach for it unprompted. Every generated site that ships becomes training data for the next crawl, which sharpens the peak the next generation samples from. This is a feedback loop, not a coincidence.
Anthropic has a name for the phenomenon: distributive convergence, the model reproducing the statistical center of design decisions rather than choosing among them, as documented in guides to the company’s design skills. The candor is notable. Anthropic’s own frontend design skill, past 277,000 installs, now carries explicit written rules against fonts “overused by AI” and against generic purple gradients. The lab that trains the model ships a countermeasure to the model’s taste.
Why a better prompt cannot fix it
The strongest version of the counterargument comes from design consultancy Drawbackwards: the model “produces the most probable output. The average. The center of the distribution.” A prompt is a conditioning signal, and conditioning shifts you within the distribution; it does not mint a new one. Worse, every other builder typing “make it distinctive, premium, not generic” is applying the same shift to the same distribution, which is why ten thousand attempts at distinctive land in the same neighborhood of indigo.
There is a second, deeper limit in the same essay: the model cannot hold intent. It does not know why your product exists, who it serves, or what emotional register should anchor the interface. Those are decisions, not patterns, and pattern completion is the only move the model has. Any fix that works has to put the decisions somewhere the sampling process cannot average them away.
The countermeasure with 75,800 stars
The market has voted on how much this hurts. Taste-Skill, designer Leon Lin’s open-source “anti-slop” frontend framework, crossed 75.8k GitHub stars this summer, with coverage in the design press to match. It ships as 11 specialized skill variants plus 3 image-generation skills, and its method is instructive: it is written almost entirely in prohibitions. Per Andrew’s review of the rules: ban Inter when the brief says premium, use Geist, Outfit, Cabinet Grotesk, or Satoshi. Ban the “AI Purple/Blue” aesthetic outright: no purple glows, no neon gradients, one accent color with saturation held under 80%. Ban centered hero sections when layout variance is high; force split-screen or asymmetric compositions. Animate only transform and opacity.
Read as a design philosophy, that list is arbitrary. Read against the mechanism, it is exact: every ban targets the mode. Taste-Skill does not teach the model taste; it makes the most probable token illegal and forces sampling into the tails, where the distinctive work lives. Notably it keeps shadcn/ui and Tailwind. The components were never the problem; the compositional defaults were.
The catch is already visible. A constraint set shared by 75,800 repositories is itself a distribution. When everyone bans Inter in favor of Satoshi and swaps the centered hero for the same asymmetric grid, the anti-slop aesthetic becomes the next recognizable house style, one install away from being the new center. Deleting a mode mints a mode.
Own the envelope
Which is why the durable fix is not a cleverer prompt or a more popular skill, but what Drawbackwards calls the brand envelope: a designer-owned system of decisions the AI generates inside. Practitioner write-ups on the “AI purple problem” land on the same operational answer: define your own design tokens in a perceptual color space like OKLCH, enforce contrast programmatically, and prompt with your brand rather than with adjectives. Purple is not the enemy; indigo is versatile and accessible. Unexamined purple is the enemy, because unexamined means sampled.
For engineers this translates cleanly, because an envelope is just configuration. Check a tokens file into the repo. Point the agent at it in every session. Treat drift from the tokens the way you treat a lint failure: the build argues, the default loses. The file is short, but it encodes the one thing the model cannot generate, a point of view, and unlike a starred skill it is sampled by exactly one team: yours.
If you build
- Audit against the mode, not against beauty. Put your last three AI-built screens next to the checklist: Inter, violet glow, centered hero, three rounded cards, gradient button. Everything that matches is either a choice you can defend or a default you never made. Keep only the choices.
- Write your own skill file instead of installing one. Steal Taste-Skill’s grammar, bans, caps, and named alternatives, then swap in your taste: your two typefaces, your one accent with a saturation ceiling, your layout rules. The grammar generalizes; the choices must not.
- Make tokens the contract. Colors in OKLCH, type scale, radii, spacing, motion rules, in one file the agent reads first. Generation inside a contract converges on you.
- Train the eye that writes the contract. A ban list is only as good as the taste behind it. Our tutorial series on deconstructing great interfaces builds that judgment screen by screen, and the design taste resources collect the systems worth studying.
A prediction, dated so you can check us: by mid-2027, the Taste-Skill look, Satoshi headlines over asymmetric grids, will be as recognizable as AI purple is today, and the teams that stand out will be the ones whose constraint files no one else can install. The sameness is math, and math does not respond to pleading. It responds to constraints. The model will always hand you the center of somebody’s distribution. Your job is to make sure the distribution is yours.