The website you are reading was built with an AI coding agent. Not a template with an AI-written headline dropped in, and not a page builder with a chatbot bolted on: the theme, the layouts, the forms, the search setup and most of the copy edits were written by an AI working in our codebase, with one of our team directing it.
We think that is worth being open about, partly because we sell software and you should know how we make ours, and partly because the honest version of the story is more useful than the hype. AI did not build this site on its own. It made building it much faster, and it made it possible to try far more ideas than we would have had time for. It also made mistakes, and catching them was a human job.
The setup
We used Claude Code, Anthropic’s coding agent, which works directly in a project’s files and terminal. One person on our team gave it instructions in plain language (“make the blog cards clickable anywhere, not just on the title”), reviewed what it did, and pushed back when it was wrong.
The site started as a quick prototype so we could settle the content and the look without worrying about where it would live. Once the design was agreed, the agent ported it, page for page, into a custom WordPress theme. WordPress was a deliberate choice: our marketing team already works in it, and our hosting is ordinary shared hosting with no build tools on the server. So the theme is plain PHP, the styling is compiled on a developer’s machine before it is uploaded, and nothing on the live server has to do anything clever.
Where AI genuinely helped
It never got bored of the details. A lot of what makes a website feel finished is tedious work nobody wants to do twice. The agent checked text colours against their backgrounds and reported the actual contrast ratios, rather than “looks fine to me”. When the logos of the businesses we work with looked unbalanced in a row, it measured how much of each image file was lettering versus empty space and sized each one so the names read at the same height. When we decided the site should not use em dashes, it rewrote around fifty sentences in context instead of doing a blind find-and-replace that would have left broken grammar behind.
It checked our claims against our own product. Early drafts of the features pages described modules our CRM does not actually have. We asked the agent to compare the marketing copy with the running product, screen by screen. Three features came off the site altogether, one was rewritten because it described something different from what the software does, and a note now sits in the code telling anyone who edits that page not to restore a claim without checking it against the product first. A website that overpromises is worse than one that undersells, and this was the single most valuable thing the AI did for us.
It made trying ideas cheap. When the home page design was not working, we did not argue about it in a meeting. We asked for six different directions, and the agent built every one of them as a separate preview page that we could open side by side and share. Then we asked for three darker takes on the one we liked. Before AI, exploring nine versions of a page was a week of design and development. Here it was an afternoon, and it meant decisions were made by looking at real pages instead of imagining them.
It handled the parts visitors never see. Pages that load the next page in the background when you hover over a link, so navigation feels instant. Images that only download when you actually open them. Search engine settings that keep preview pages out of Google. Structured data that tells search engines who we are. None of it is glamorous, and all of it would normally get cut for time.
Where it got things wrong
AI makes mistakes quickly and confidently, which is a dangerous combination if nobody is checking. A few real examples from this build:
- Asked to remove some old code, it deleted a whole file. The file also contained a small function the site header depended on, so the site stopped loading. It noticed straight away because it tested the change, and fixed it, but on a live site that would have been an outage.
- It proposed cropping blog images to a standard widescreen shape. When it measured our actual screenshots, they turned out to be wider than that, so the crop would have made them taller and cut off the sides. It caught this itself, but only because it measured instead of assuming.
- One of its edits introduced a stray space before a full stop in a sentence. Small, but exactly the kind of thing that makes a site look careless.
The pattern was consistent: the agent was reliable when it verified its work against something real, and unreliable when it assumed. So we set things up to make verification easy. Every change was tested on a private copy of the site running on a laptop before it went anywhere near the live one, and the agent was expected to load the page and confirm the change rather than declare it done.
What people still did
The AI wrote most of the code. People made every decision that mattered.
We decided what the site should say about us, and we supplied the facts: our registrations, our memberships, our contact details, our certificates. We rejected a lot of the AI’s work, sometimes several times in a row. The credentials section on the home page went through four layouts before we settled on one, and the home page background went through more versions than we would like to admit. Taste, judgement about what our customers care about, and the final say on anything that goes live all stayed with us. So did responsibility: when you visit this site, everything on it is something a person at Tech Bridge Consultancy chose to put there.
What we took away from it
AI did not replace the work of building a website. It changed where the effort goes. Less time typing, much more time deciding, reviewing and testing. A team that knows what it wants, and can tell good work from bad, gets enormous leverage from these tools. A team that cannot will just produce more mistakes faster.
That is also how we think about AI in our own product, and it is why we wrote two companion pieces: whether AI-accelerated companies still need a CRM, and why you might use ours rather than build your own with AI.
If you would like to see what CherryBerry looks like configured around the way your team works, book a walkthrough. A person will reply.