What a prompt gives you
One good-enough page, and no way to get the next one to match it. The brand lives in the prompt, which means it lives nowhere — every edit is a fresh negotiation with a model that has forgotten what you agreed.
Project
The control layer for AI-built websites
An LLM will build you a website in fifteen seconds. What it will not give you is control over what it built — on-brand, on-strategy, and provably so.
Generation is commoditized.
Control isn't.
The problem
Ask any model for a marketing site and you get something fine and forgettable — stock phrasing, a layout you have seen a hundred times, sections that do not quite fit the business. The failure is not the model's taste. It is that nothing holds the model to anything. Ask again tomorrow and you get a different site, off in a different direction.
One good-enough page, and no way to get the next one to match it. The brand lives in the prompt, which means it lives nowhere — every edit is a fresh negotiation with a model that has forgotten what you agreed.
A brand that exists as files — a voice, a visual style, a page shape — that every page is generated from and checked against. A page that would drift off-brand fails a check and never ships.
How it works
Auxon is not a website builder and it never calls a model as its own judgment. It is an alignment layer — it holds the definition of the brand, checks the work against it, and refuses what drifts. That division is the whole point: a system that both writes the copy and grades it is not checking anything.
A voice pack (how it sounds), a style pack (how it looks and moves), and a shape (what each page is). Real CSS, real rules, versioned and committed — not a paragraph of instructions the model may or may not honour on the next turn.
Your coding agent writes every line, with the brand in front of it every single time rather than in its memory from an hour ago. Auxon ships an operating manual that teaches the agent to drive it, including what it must never do.
Around sixty automatic checks run after every change — banned phrases, required sections, mobile, accessibility, real images, link integrity. Some are literal string checks with no model judgment at all. What fails does not ship.
Where a brand comes from
You rarely start from nothing. The brand is on a live site, in a screenshot someone sent you, or in a founder walking you through a competitor's homepage on a call. Auxon turns any of those into a reusable pack.
The richest source. Auxon also drives a real browser against the page to read its motion off the live DOM — transitions, keyframes, hover and focus behaviour, scroll reveals, even the drawers and modals a screenshot can never show you.
For a competitor behind a login, a design comp, a deck slide. Pixel sampling gives measured values instead of guessed ones, and a bundle of screens lets repeated motifs surface the actual design system.
Frames sampled on a timeline and the narration transcribed, so "notice how the cards lift on hover" lands in the pack as design intent, tied to the frame that shows it.
What it does
A site has a life after launch day. Most tools stop at the first build, which is where the real work starts.
Compose voice, style, and shape; emit the pages; prove they conform.
Scan for drift, make a surgical edit, and prove nothing else moved.
Stand up on-brand A/B variants, read honest measurement, promote a winner safely. It reports the real numbers and flags when a test is not decision-grade yet — it never picks a winner for you.
The business changed. Cascade that through the brand source and re-author only the pages it actually affects.
The part that matters commercially
Most brand tooling requires you to rebuild your site inside it, which is why most of it never gets adopted. Auxon captures the brand off a live WordPress, Squarespace, Webflow, or bespoke site and binds the site to it in place. You can judge any page or any proposed edit against that brand, keep it on-brand as it changes, and run conversion experiments on it through a single embed. No migration, no platform to move onto.
For whoever writes
Describing an edit to an LLM is slower than typing it. So the words have their own layer, where every field has an exact address and edits apply deterministically — no model in the loop, brand rules still enforced.
Navigate to any page, edit its words, then import and verify from the same screen.
Click any element on the live site, type the new text, review a queue, apply.
Export every word to Markdown for a copywriter, then import it back.
Status
Auxon is in production on my own sites and client work, distributed as a self-contained binary. The engine inside it is called sitekit — named rather than hidden, because the point is that the outcome is engineered instead of prompted. The documentation is public and current.