Last time we wrote about what made us rebuild the website. This time it is how we built it.
The short version: we built it in conversation with an AI. We did not go to a web agency.
How it actually went
The method was simple — talking it through in a chat window.
It started with “I want to build a site for a restoration company”, and from there the structure, the writing, the design and the choice of photos were decided one at a time by talking them over. If we said “that is not right”, it changed. If we asked “why did you do it that way”, we got a reason.
The difference from commissioning an agency is that there is no waiting. You do not send a revision request and wait a few days. In exchange, every decision stays on your side.
What we handed to the AI
- First drafts of the text
- Proposed page structures
- Design and code
- Research into what other firms in our trade have on their sites
- Measuring load speed and how much text there was
That last one was a help, because it is tedious work for a person. We said “the home page feels wordy”, and it went and measured it in a browser and came back with “2,481 characters, about 8 screens’ worth”. An argument about impressions becomes an argument about numbers.
What we did not hand over
Deciding whether something was true.
That was the whole of it. An AI will write plausible text all day long. And awkwardly, what it writes looks true.
Lines we actually deleted
Some of the text we read back and deleted. A few examples.
“We keep to the schedule. You will have a programme of works before we start.”
A plausible sentence. But restoration is a job where you strip the wall and then find the substrate has gone. There will be days we cannot keep to. The moment you write it, it is a promise waiting to be broken. Cut.
“We quote materials and labour separately.”
This did not match how we actually work either. We can provide a breakdown on request, but we do not always do it that way. “We can” and “we always do” are different things, so we reworded it.
A list of the wards we cover
The ward names the AI lined up were ones nobody had checked against reality. A textbook plausible falsehood. We corrected it to the wards we actually work in.
“◯ years in business, ◯ jobs completed”
We started writing this and stopped. You cannot publish a number you have not counted.
What we took from it
An AI can write text that sounds like fact without knowing any of the facts.
And people, reading tidy well-formed prose, assume it is probably right. That was the dangerous part.
So our job was not writing but deleting. Going through what came out line by line and asking “is this true?” and “can we keep to it?” If we could not check it, it went.
The same thing happens when you commission an agency, in fact. The writer does not know your company, so they have to fill the gaps with generalities about the trade. The only difference is that an AI does it faster, in greater volume, and more plausibly.
This post was drafted by an AI too
For the avoidance of doubt: OS Notes posts are drafted by an AI and checked by a member of staff before publishing. We put a note to that effect at the bottom of every post.
It did not seem like something to hide. If anything, saying how far the AI was involved makes it easier for the reader to judge.
Next time: how much of the writing we cut.
In this series
- The phone was the only way in
- Building it alongside an AI (this post)
- Cut, cut, and cut again
- Loose ends outstanding before launch