We let AI write our whole website — with one rule: don't change how it looks
Last month I gave an AI a job I wouldn't hand a brand-new hire in their first week: rebuild our company website.
One rule: it couldn't change how any page looked. Not the colours, not the text, nothing a visitor would notice. The site already worked and was bringing in leads, so it had to look identical afterwards. But underneath, everything had to change — from a pile of hand-built pages into a proper system we could actually publish blog posts to.
Rebuild everything. Change nothing anyone can see. Strange rule. Also exactly why it was worth doing.
What most people get wrong
When people hear "we let AI write our code," they assume the AI is the scary part — the one that does something dumb and breaks the site.
That's not what worried me.
What worried me was simpler: if I let anyone touch a live website — an AI, a new developer, anyone — how do I actually know they didn't break it? You can't just hope. You need something that checks, automatically.
So before the AI wrote a single line, we did the least exciting thing possible. We went through our own code first and wrote down exactly what "working" means — every rule, everything that can't break.
That list was the whole secret. An AI does exactly what you tell it, as well as you tell it. Give it a fuzzy goal, you get fuzzy work. Give it a precise checklist, it's brilliant. The checklist — not the AI — is what keeps you safe.
The safety net we built first
Then we set traps. Small automatic checks that scream if anything goes wrong:
A test that screenshots the site and fails if even one pixel moves.
A check that fails the build if a page quietly disappears.
A setting that refuses to even start the app if a password is wrong, instead of pretending it's fine and breaking later.
Each one is a spot the AI could have quietly broken something. And couldn't — the trap was already there, waiting.
(We turned that whole safety net into a simple checklist you can copy. Link at the end.)
For a while it worked beautifully. The AI built the blog, the authors, the tags, all of it, under a site that never flinched. I relaxed.
That was the mistake.
Because a few days after it went live, every one of our old pages quietly ignored the new work. No error. No warning. Everything looked green and fine. And it took us embarrassingly long to figure out why.