Building a website with AI: what worked and what didn't

One project, several AI sessions

I built my first website as a schoolboy, at a time when you still typed HTML by hand. This one I rebuilt with AI. The main tool was Claude Code, an AI programming tool from Anthropic that works through tasks independently in several steps. The result first: the AI works faster than any developer I know. But it needs clear rules, and in the end a human has to check what it delivers.

What surprised me was how well working with several sessions in parallel went. One session built the admin tool, my own lean editorial system in PHP. Another developed the blog strategy and the writing rules. The sessions coordinated among themselves who was editing which files, so they would not overwrite each other. That was less programming than project management, with a team that never gets tired.

Speaking instead of typing

A large part of my requests came in by voice. Often several in quick succession, even in the middle of ongoing work. While testing I noticed something, I said it out loud, and the AI took it into its current work. For someone who has spent twenty years translating every wish into code himself, that is a different way of working. I describe, the AI builds, I look at the result.

The admin tool itself shows how far that goes. It writes posts by AI from bullet points or takes over my own texts. It generates photos by AI, sets the details for Google and prepares the LinkedIn preview. For testing it runs locally with a small database, in operation with the database at the hosting provider. For an editorial system used by only one person, that would once have been weeks of work.

The first post was nonsense

The first blog post the AI wrote was about hourly fees in law firms. It sounded elegant. It was deeply pompous, barely comprehensible, a nonsense text. Claims without evidence, images instead of facts, plus first-person statements I had never made. Had I published it unread, my name would have been underneath.

Out of that came a binding writing guide: invent nothing, no stock phrases, no direct address, no gendered forms, short sections with short subheadings, always a conclusion box, texts factual but entertaining. Since then the drafts have been usable. The lesson: the AI writes only as well as the rules it is given. Without rules it delivers the average of the internet, and that average is pompous.

What went wrong

For the blog, a topic list with over 200 topics was created, including researched court decisions from employment law, data protection and AI. AI agents write the drafts in parallel, check their sources and flag uncertain points. I read and approve. That sounds like a clean machine. Three things went wrong all the same.

First, the initial drafts ended up in the local test database instead of the live system. The AI had done nothing wrong, I simply had not said where the texts should go. The solution was an import button in the admin tool. Second, the AI had initially misunderstood a rule on judgments and had to correct it. Third, two posts on the same judgment were once produced, because two agents had each picked up the same topic. None of this was serious. All of it would have been published without checks.

Small things that make the difference

For me, what decides whether a piece of software is good is not the technology but whether people like using it. That applies even to a tool that only I operate. So I insisted on details an AI would never come up with on its own. The contents box next to the post must not scroll, so no more than seven subheadings. Publishing works via a toggle, with no extra saving. Employment law, data protection and AI got their own topic areas.

With the images I was strict: real photos, bright, no cartoons, no text in the image. And one decision that matters to me: the posts carry no AI notice, because I have read and approved every sentence. The AI images carry one, because there nobody checks anything. My AI assistant Eora on the website is, incidentally, currently a demonstration. As a real AI she is planned for a later stage.

What I would advise others

Anyone having a website or an internal tool built with AI should plan for three things. First, rules before the start: what may the AI do, what not, what should the result look like? Every rule that is missing it fills with an assumption, and the assumption looks finished. Second, time for checking. The effort does not disappear, it moves from building to reading. Anyone who pockets the time saved in full will publish mistakes.

Third, a person who knows the day-to-day work. The AI builds what is described. Whether the contents box scrolls or the publish button takes one click too many is noticed only by whoever uses the software later. That person cannot be replaced by computing power. That is also the answer to the question of whether I would rather have programmed the site myself. No. But I would not have left it to anyone who does not look closely.