Why the AI in nqrah never writes code

When people hear that nqrah’s Atelier designs a store from a description, they usually imagine a model writing HTML and CSS. It doesn’t, and that is a deliberate decision.

#What Atelier actually does

A merchant describes their brand in a sentence, or answers a few short questions. Atelier then builds the store, a page or a single section by choosing and filling sections from a library of 31 tested section types: heroes, collection grids, product details, reviews, FAQs, checkout blocks and so on. The model’s output is structured data: which section, in what order, with which copy, images and settings. The platform validates that data against a schema and renders it with components we wrote and tested.

So the model makes design decisions and writes the words. It never writes the code that runs in the customer’s browser.

#Why not let it write code?

#Checkout is sacred

A store exists to take orders. Free-form generated code can break a layout on one phone, hide a button behind a banner, or interfere with the cart. With a library of tested sections, there is no output the model can produce that breaks checkout, because checkout is not something it can touch. The AI doesn’t edit the cart, prices, inventory or payments at all.

#Quality has a floor

Every section in the library is accessible, responsive and right-to-left ready. When the model composes from them, the worst case is a store that is less inspired than it could be, not one that is broken.

#Changes are reviewable

Because the output is data, every change Atelier proposes can be shown as a preview before it is applied. Nothing is published until the merchant decides, and every version is saved so they can undo. Reviewing a diff of “hero changed from split to centered, headline rewritten” is something a shop owner can do. Reviewing generated CSS is not.

#It is cheaper and faster

Generating a structured plan uses far fewer tokens than generating a page of markup, and the result renders instantly with code we already ship. That keeps AI cost predictable, which matters when every store has a budget.

#The trade-off

Constraints have a price. Atelier can’t invent a layout that isn’t in the library. When merchants keep asking for something we don’t have, that is a signal to design and test a new section type, once, for everyone. The library grows deliberately instead of the platform accumulating thousands of slightly different generated pages.

#The general lesson

For a lot of AI features, the best interface between the model and the product is a schema, not a code editor. Let the model do what it is good at: understanding intent, choosing among options and writing copy. Let tested software do what it is good at: rendering, validation and keeping promises like “checkout always works”.

Merchants still get a store that looks like their brand. They just never have to wonder whether the AI broke it.