The problem
This site is the clearest example we can show you. We built it in phases: the copy was agreed
before any page existed, and a working preview was there to review before anything went live.
A project with us follows the same steps: decide what it should say, build a piece, look at
it, move on. Here, a set of automated checks also runs underneath every step.
The data
There's no dataset behind it. The inputs are a written specification and a facts file that
records what is true about the business, and nothing reaches a page unless it's in that file.
That single rule is why you won't find borrowed logos, invented testimonials or made-up
numbers anywhere on this site.
How we approached it
A rule written in a document gets forgotten. So every rule we agreed became a test in the
build, and breaking one fails the build.
On your own project, that would mean:
- No page can publish a price by accident
- No unfinished page can be found by Google
- Nothing private can be published
- No page can quietly lose the label that says what it is
- No claim can appear that the facts file does not support
None of that depends on someone remembering it at four o'clock on a Friday. The checks run on
every change, and a failure stops the change.
How it is built
Every page is finished HTML before a browser asks for it, so there's nothing to wait for and
nothing that can go wrong while it loads. A written rule set decides what may be claimed, and
the checks described above enforce it on every change.
What the AI did, and what a person decided
A person on our team decided the positioning, what the site may and may not claim, the pricing
policy, the voice and every word that ships. The AI assistant wrote the code, drafted copy for
approval and built the checks.
Anything the assistant proposed that nobody had asked for was marked as a proposal, in the
open. We had it remove two pieces, including one it had argued for.
What we can show
The most useful thing we found was a problem in our own quality checks. One of them had been
reporting success while checking nothing at all: a setting meant to skip a couple of links was
skipping every link on the site.
We fixed it, then broke a link on purpose to watch the check fail. Since then,
every check gets tested against a real mistake before we trust it, because a
green tick that has never seen a failure proves nothing.
What it demonstrates
On your project, the rules that matter to you get written down and turned into checks, and the
checks get tested too. You see a preview before anything is final, and the things you said
mattered still hold after the enthusiasm wears off.