How to build with AI without losing quality
AI output drifts once you ask for more than a few things. Set plain rules first, check every output against them, and keep the final yes.
On this page 5 min read
Key takeaways
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
You tried AI on your product, and the first few results looked great. Then you asked for fifty more. Now the buttons don't match, the icons look like they came from five different sets, and you spend your evenings fixing things one by one.
The fix is to set the rules before the AI makes anything, then check every output against them. AI handles the volume. A person sets the rules and decides what ships.
Here's how that split works, and how to use it on your own product.
Why does AI work get messy after the first few outputs?
Short answer: each output is made on its own, so small differences pile up until nothing matches.
Ask for one icon and you get a good icon. Ask for a second and that one is good too. Put them side by side and the lines are a bit thicker on one and the corners a bit rounder on the other. Neither is wrong. They just weren't made to the same rules.
With ten outputs you can fix that by eye. With hundreds you can't. And it's the same story with landing page sections, product screens and emails. Each one looks fine alone. Together they look like five people made them.
Here's the difference between two ways of asking:
- Before: "Draw me a settings icon." You get a settings icon in whatever style came out that time.
- After: "Draw a settings icon on a 24px grid with a 1.5px stroke, round corners and 2px of padding inside the grid." You get an icon that can sit next to the others.
The second request is longer and less fun to write. It's also the one that still works at number 500.
There's a second reason it drifts. When you leave a gap, the tool fills it with its most common default. That's why so many AI-built sites look alike. I cover the visible signs in why does my AI app look generic.
What rules should I set before the AI starts?
Short answer: the few things that must be the same in every output, written so plainly that a script could check them.
Say you run a small online shop and want AI to help build product pages and an icon set. A first rule list might be:
- One grid. Every icon sits on the same 24px grid.
- One stroke. Every line is 1.5px, with round ends.
- One set of colours. Text, background and one accent. Nothing else.
- One type scale. Five heading and body sizes, used everywhere.
- One spacing scale. Gaps of 4, 8, 16, 24, 32 or 48px, and nothing in between.
A good test for a rule is whether you can answer it with yes or no. "Make it look clean" fails that test. You and the AI will both read it differently every time. "Every icon sits on the same 24px grid" passes. It's either true or it isn't.
Keep the list short. Five rules that are always followed do more than thirty that are followed sometimes.
Write the rules down in a file the tool reads every time, not in a chat message you'll forget to paste. Design teams call these values design tokens: named colours, sizes and spacing that every screen pulls from. Once the values live in one place, the AI picks from your list instead of inventing a new shade of grey.
What should AI do, and what should a person decide?
Short answer: let AI help with the volume. Keep the rules and the final yes for a person.
AI is good at the repetitive part. Drawing the fortieth variation, filling in names and search tags, turning one approved layout into ten pages, writing the code for a component you've already designed.
The calls a person should make:
- The rules. Grid, colours, type, spacing. AI can suggest options, but someone picks.
- The naming. What each thing is called, so people can find it. An icon you can't find is an icon you can't use.
- Exceptions. Brand logos, special pages and anything that is allowed to break a rule, and why.
- What's missing. Look at the whole set and ask what a user will need that isn't there yet.
- The final yes. Someone reviews what goes out.
My default split is simple: the tool proposes and a person decides. If nobody owns the final yes, the tool's defaults end up owning it.
How do I check hundreds of outputs without losing my weekends?
Short answer: let a script catch the rule breaks, and save your eyes for judgment.
A lint step is an automatic check. It runs on every file and flags anything that breaks a written rule. The name comes from lint, a code checker Stephen C. Johnson wrote at Bell Labs in 1978. For design work, a check can look for colours outside your palette, spacing values outside your scale, or icon strokes that aren't 1.5px. Stylelint can do some of this for CSS.
This changes what review is for.
- Without the check: review means hunting for mistakes. Is this line the right thickness? Does this one sit on the grid? Multiply that by hundreds of files and it never ends.
- With the check: the script answers those questions. You only look at what a script can't judge. Does this icon read as the thing it's named after? Would a customer understand this page?
Review catches problems one at a time. A rule with a check stops the same problem from showing up again.
If you can't write a script, a checklist works. Run it on every output, every time, and answer each line yes or no.
Does this work if I'm not making icons?
Short answer: yes. It works for anything you make in volume.
Say you're using AI to build pages for your site. Rules you could set first:
- One main button per page, and it always looks the same.
- One set of heading sizes, used everywhere.
- One set of spacing steps, and nothing in between.
Then check each new page against that list before it goes live. A checklist you run by hand is fine to start with. The point is that the check happens every time, and the answer is yes or no.
The same goes for emails, social graphics and product screens. Pick the few things that must never change, write them so a stranger could check them, and check them.
And if you're hiring someone who builds with AI, ask two questions. What are your rules? How do you check them? If the answer is "I look it over", expect the fifty-icon problem from the top of this post.
Want a site or app built fast that still looks like one person made it? Book a call and tell me what you're making. I reply within a day.
Questions
-
It can produce a lot, fast. It doesn't decide what good looks like for your product. Someone still has to set the rules and say no to the outputs that break them.
-
It's an automatic check that runs on every file and flags anything that breaks a written rule. Programmers have used linters since 1978, and the same idea works for design rules.
-
Five or so. A short list that is always followed does more than a long one that is followed sometimes.
Read next
View all posts
Best text for a call to action button (not Submit)
The best text for a call to action button names what happens after the click. See rewrites of Submit, Click here and Learn more.
Consistent icon style in UI: why mixed sets look cheap
Mixed icon sets make a UI look stitched together. How to spot it, why emoji fail as icons, and how to pick one family and weight.
How to design a website without content (no lorem ipsum)
Copy not written yet? You can still start designing. Use rough draft copy instead of lorem ipsum. Here's how, step by step.
Have a project in mind?
Book a short call. We'll talk through what you're building, and I'll tell you how I can help.
- Length
- 15 minutes
- Where
- Cal Video
- Time zone
- Metro Manila