Content without a classic CMS

This site has no CMS: I say what I want and an AI builds it. For you it is an offer you can sell for a flat fee.

01

The moment

The page you are reading has no admin panel. There is no dashboard and no menus to click through. The text lives in plain files, and I told an AI in plain sentences what to create, change and publish. This text reached the site that way too.

The question behind it was simple: can an AI do the work a content management system usually does for a small site? Not as a toy, but day to day, with real changes, real mistakes and a site that is public.

For you this counts because you probably spend part of your working time commissioning, aligning or clicking together changes to websites in a CMS. And because a website an agent looks after is an offer you can sell for a flat fee.

02

What went wrong

What went wrong was my mistake, not the AI’s. When something is not clearly defined up front, the AI invents it. An invented section usually lands just outside the brand. The colours are almost right, the spacing is almost right, the tone is almost right. With a brand, almost is the opposite of right.

The mistake behind it: at first I asked for sections without showing first how they should look. The AI did exactly what I said, and that was too little. A vague request gets a vague result. You know that from every agency that gets a bad brief. How many sections I had rebuilt, I will add.

03

What I tried

The rule that fixes it is one sentence: design the section in Figma first, exactly as it should look, and only then build it. Skip that step and you get something plausible that does not belong to the brand. Keep it and the drift disappears. On top of that there is a test environment: changes land on a copy first and only go to the real site after a look at them.

Everything technical on this site runs on Google Cloud. With a written guide I taught the AI which building blocks of the platform to use for what. The experiment is still running, and the details are on the experiment page.

04

What I know now

What I know now: structure changed the most. The better a site is organised up front, with named sections and a layout that holds together, the better the AI understands what is wanted. Once that scaffolding stands, a request like add an article with this content is enough. The AI knows where it goes.

Design comes first, every time. And the AI works better when it has the rules of the site in writing in front of it: which fonts, which spacing, which building blocks exist and which do not. A short brand description next to the code is for the AI what a style guide is for an agency.

Then the money. If a business pays an agency today for changes to its website and that agency still works by hand, it pays the price of hours for minutes of work. The better deal is one where the client sends what they want, a voice message if need be, and the provider’s job is making sure the AI is used the right way.

That is where your offer is. Many small businesses have a website that goes stale, because every change is a little project. If you shorten that path to one sentence and charge a fixed monthly price for it, you are not selling hours but a website that stays current. And because the website is where enquiries are born, it belongs right in your niche: data and agents for marketing.

The limit is where many people work on content at the same time, need approvals and manage permissions. Then a CMS is more than an interface. For the small site of a small business it is often just ballast today.

05

Open question

What is open is how well this holds over years. A site built only from requests can become inconsistent over time if nobody watches the whole. I am watching that on this site and will record here what I see.

06

Changelog

29 October 2026
First version. Nothing here is rewritten silently: every later change gets its own entry in this changelog.

Next episode: What still sells when AI replaces competence