How do you stop AI from producing generic slop?

AI has a look, and it's bad. Cards that change height depending on how long the text is. Items in the bottom right drifting around. The card with one colored edge on the left or top and nothing around the rest. Readers notice, and so do I.

"It'll make cards just different sizes based on how long the text is, which is obviously horrible design."

What are the tells that AI made it?

The one-sided card stripe is the big one. A colored line on one side of a card and nothing on the other three is the em dash of design. It's banned in my system. Put the same strong outline on all four sides instead.

• Cream on cream. Cards in almost the same color as the background. I can deal with cream as long as everything isn't cream.
• Cards that resize with the text. Same content type, same card geometry.
• No images where images are obvious. A city page with no city photo is an awful choice. There are obvious overview images for every city.
• Too much text. Too many bullet points that don't seem relevant, and it reads like a boring local business page nobody wants to look at.
• Dark mode by default. On a media kit I was fighting, dark mode was part of the problem.

Telling the AI once doesn't fix it. I've had an agent say it added the rule to its notes and then do the same thing again. Rules have to live in the design system and the instruction files, and a reviewer has to check for them.

Why does AI design look so generic?

Because when you don't give it a standard, it uses its defaults, and everyone else's AI uses the same defaults. When you start from zero, it gives you crap. The fix is to give it something better to follow, every time.

My design order, from most to least authority:

1. A specific approved reference for this exact thing
2. The current brand system
3. An approved template
4. My durable taste (things I always want or never want)
5. The model's generic preference, last

Approved systems and assets outrank improvisation. Inventing a plain new look from scratch is not acceptable. Either reuse the design system I already have or use a dedicated design tool and do something much better.

What bar should the design hit?

A-minus. I don't need an A or an A+. The AI drafts I've graded landed around B-minus at best, and that gap is all specifics: spacing, real photos, the right palette.

The test I give my AI: would a normal small business owner look at this and think it cost $20,000 from a designer, minimum? For client work, I want something that would impress an eight-figure company. Your homepage gets the highest bar of all, because everybody lands there and it has to look like a million-dollar local media brand.

Competent is failure. Safe is failure. I'd rather see something too bold that I pull back than another decent SaaS page. And the AI has to judge its work the way people see it:

"If all you're doing is looking at the code of this stuff, you've got to stop. You have to look at it with eyeballs to understand how it visually looks to humans."

How do I build a design system the AI will follow?

I moved all my websites onto one design system, the one Claude came up with for a news site I built. Landing pages, my software, my newsletters: just use that system. Now they all look the same, and I don't have to think about it anymore. Mostly. The system is two things:

• Markdown instruction files the AI reads: what to do, what's banned (like one-sided colored card borders), consistent card geometry
• CSS design tokens: the actual colors, fonts and spacing

The app shell can stay consistent and simple. The content that sells, like slides and presentation graphics, should match the colors but be far more detailed and unique than the shell. I'd like to productize the system: a business drops in its logo, colors and domain, and every document and page comes out on-brand.

How do you get a good design out of an AI builder?

1. Decide the strategy and exact copy first. A media kit that had the AI invent the content and the design together came out generic. Lock the words, then ask only for different designs of the same thing.
2. Don't overload website builders. My habit of having Claude write a long, detailed prompt for another AI backfired with website builders. They're good at taking the basics and making them pretty, and a super-specific prompt makes them try too hard. After about 15 tries on four different systems, I asked Claude for just the copy and got a great result in three or four tries in Google AI Studio.
3. Screenshot sites you like. Hand them to the builder. I forget to do this, and I should do it every time.
4. Ask for at least three concepts. For a redesign, I screenshot the current software, explain who uses it, and ask for three modern directions that aren't so unusual people struggle to use them. Then I have Claude write the prompt for a builder like Lovable to make the chosen one. A lot of my best design prototypes came out of Gemini.
5. Keep what works, redo the rest. When cards keep coming out wrong, I keep the parts that work, like the toggles, titles and buttons, and ask for three new takes on the card itself, the fonts and the colors.

What about documents and Notion pages?

• Human-facing docs should be readable. I want a Google Doc or Word style document. Very beautiful, easy to read. I'm not going to read a raw markdown file.
• No walls of text. When I rebuilt my CRM in Notion, telling the model to avoid walls of text and use visuals made a big difference. The top model does design well. A smaller one doesn't.
• Inspect the rendered output. The AI should look at the finished page, not just the code, before showing it to anyone.

How do I stop slop in writing?

• Banned words. My reviewer agent checks for banned words and terms before anything reaches me.
• The out-of-state test. For local content, if a story's insight could have been written from out of state, it isn't worth publishing under a local brand. The value isn't that a Chick-fil-A is opening. It's knowing that intersection was already backed up. AI can't know that.
• Human on the front end. Against a flood of AI content, the differentiator is that a real person is behind the brand. AI goes on the grunt work. Readers and advertisers get a human.

The last line of defense is a reviewer agent with two or three passes before anything reaches you: should an AI reviewer check work first. Put the standards in your agent instruction files so every build starts from them.