Stop Reviewing AI Design. Encode Your Brand Rules Once and Never Review Again.

There's a loop everyone falls into with AI-designed websites. You ask Claude to build a landing page. It comes back looking... fine. Wrong headline voice. Wrong shade of blue. A form that asks for a phone number you'd never collect. You write a paragraph of feedback. It fixes three things and breaks two. You do this four times, and the page that took ninety seconds to generate took an hour to correct.

The problem isn't the model's taste. It's that you're treating design review as a conversation when it should be a configuration. Everything you typed into that feedback loop — every "no, warmer," every "we don't say 'solutions'" — is a brand rule you already knew. It was sitting in your head the whole time. Encode it once, up front, and every page comes out right the first time.

Design quality debates are a distraction

There's a whole genre of content now that ranks AI design output — show ten site builders the same brief, screenshot the results, crown a winner. It's entertaining, and occasionally useful, but if you're publishing with an LLM it answers the wrong question. You don't control the average output of a model. You control the instructions you hand it.

We've covered this from other angles already: keeping your cheap domain and letting Claude and ChatGPT do the hosting work, and wiring a model into your own site instead of renting a builder. The stack matters. But the layer almost nobody talks about is the input layer — the standing instructions that make "generated by an LLM" and "looks like it came from your brand team" the same sentence.

The two files that replace design review

Everything I'm about to show you lives in two artifacts you write one time:

1. A brand rules file (brand.md) — the facts and constraints. Voice, banned words, color values, layout rules, conversion requirements. Claude reads this before every generation.

2. A system prompt — the operational instruction. "You are the site generator for [company]. Read brand.md before producing anything. Never deviate."

If you're using the Claude or ChatGPT setup we walked through before, this file gets pasted into your project's custom instructions or saved alongside your publishing endpoint, so it's loaded every single time without you lifting a finger.

What goes in the brand rules file

Here's the actual structure. Steal it wholesale and fill in your details.

# Brand Rules — [Company]

## Voice
- Direct, plain sentences. No hype.
- We talk to one person, not "teams" or "users."
- Banned words: solutions, seamless, elevate, unlock,
  revolutionize, cutting-edge, empower.
- We never use exclamation marks in body copy.
- First person singular ("we" only if the site is a team).

## Visual rules
- Colors: bg #FFFFFF, text #111827, accent #2563EB,
  error #DC2626. No other colors without permission.
- Type: one sans-serif family, max two weights per page.
- Headlines: 48–72px desktop, 32–40px mobile.
- Max content width: 640px for reading, 1100px for page shells.
- No drop shadows, no gradients, no rounded corners over 8px.

## Conversion constraints
- One CTA per page. One. The button text states the action
  ("Start free trial"), never "Learn more" or "Submit."
- Hero must contain: what it is, who it's for, the CTA. Nothing else.
- Forms: email only unless a field is legally required.
- Every claim gets a source or gets cut.
- Social proof, if any, goes below the fold. Never fake logos.

## Structure
- Every landing page: hero, proof, how it works (3 steps), FAQ, CTA.
- Mobile first. If it doesn't survive 375px, it doesn't ship.
- Pages load as plain HTML + one CSS file. No frameworks unless asked.

That last line matters if you're publishing to static hosting — we covered why plain HTML wins in the full setup from domain to published post.

The system prompt

The rules file is the constitution. The system prompt is the enforcement:

You are the page generator for [Company]. Before producing
ANY page, copy, or design decision, read brand.md and treat
it as non-negotiable.

- Output complete, deployable HTML. No placeholders, no
  "add your logo here."
- If brand.md conflicts with my request, say so and follow
  brand.md unless I explicitly override.
- If brand.md doesn't cover a decision, make the most
  conservative choice and flag it in one line after the code.
- Never invent statistics, testimonials, or customer names.
- After the code, list any brand.md rules you were unsure
  how to apply, in one line each.

That last section is the quiet superpower. Instead of you hunting for violations, the model self-reports its judgment calls. Your review drops from "read every pixel" to "read five flagged lines."

What changes in practice

Before the rules file, a typical page cycle looked like: generate, read, list eight problems, re-prompt, re-read. Now it's: generate, skim, ship. The model isn't guessing at your taste anymore, because your taste is a document.

Two things happen that surprise people:

The pages get more consistent than your hand-built ones. A human freelancer drifts. A rules file doesn't. Page twenty looks like page one, because they're both being generated against the same constraints.

The model catches its own scope creep. Ask for a pricing page and it'll flag, unprompted, "brand.md says one CTA per page — this has three, so I picked the primary." You've turned a design conversation into a compliance check.

Maintaining it

The file isn't carved in stone. When you find yourself giving the same piece of feedback twice, that's the signal: stop giving it as feedback, add it to brand.md. The file should grow for a few weeks and then flatten out — most brands stabilize at thirty to fifty rules.

The one habit worth keeping: when a page does go wrong, don't fix the page. Fix the rule that let it happen. A page-level fix means the same mistake shows up in the next generation. A rule-level fix means it can't.

That's the whole shift. Everyone else is debating whether AI design is "good enough." You've moved the question to the layer you actually control — and once the rules are written, every page the model ever generates is already on-brand before you look at it.