blog
Your AI has no memory of any real page
Every model reaches for the same defaults because it has read the internet's average, not the specific work. A library of 577 studied products gives it named moves to steal and named tells to refuse — before it writes a line.
The same page, every time
Ask a model to build a landing page and watch what arrives. A centered hero. A soft gradient behind it. Three feature cards with icons. A testimonial row. The same generous border radius on everything. It is competent, and it is the same page your competitor got when they asked.
That sameness is not a bug in the model. It is what averaging looks like. A model trained on the whole web has read more pages than any person, which means it has absorbed the most common patterns hardest — and the most common patterns are the ones everybody already has. What it has not done is look closely at any single page and notice what that page did on purpose.
What a card does that a prompt cannot
A prompt is a wish. A card is a reading. Each card in the library names a real product or a real technique and writes down two things: what to steal, and what to avoid. Not "use good typography" — the specific move, on the specific page it came from, with the reason it works there.
That distinction matters because the useful part of design knowledge is not the pattern, it is the judgment around it. Why this page can carry a heavy display face and yours cannot. Why that storefront's empty space reads as confidence and this one's reads as padding. A model that has only seen patterns cannot make that call, because it never saw the decision.
One direction, with hard rules
Most AI builds drift because they are chosen by committee of vibes. A little of this reference, a little of that. By the end the page has four type scales and no point of view.
The fix is a constraint you can fail. Before anything is built, one direction gets committed to: a palette with a limit, contrast floors that are numbers rather than adjectives, a motion budget that says how much movement is allowed. Then the build either ships against that direction or it does not. Direction is not a mood board. It is a rule with a test attached.
Grade the pixels, not the vibe
The last step is the one most builds skip. Once the page is rendered, a script measures it: the actual contrast ratios of the actual text on the actual background, whether the layout logic holds, and whether any of the known tells are present in the markup. It returns PASS, or it returns the exact list of what to fix.
Measured beats admired. A contrast ratio is not an opinion, and a page that fails the floor cannot be talked into looking good. The agent that built the page does not grade it — writer is not the reviewer — because builders are reliably generous with their own work.
What we refuse to fake
The brain is free, there is no paid tier, and there is no dashboards-worth of numbers to impress you with. The library is 577 cards: 497 studies and 80 technique cards, open to read, with the JSON served to any agent that asks. The honest state of everything else — what is open, what is a waitlist, what has never billed — is on the states page, and it says nothing is live until it is.
If your AI keeps building the same page, it is not because the model is bad. It is because nobody showed it any real work first.