KI und WordPress WP Meetup Hamburg #140

Chapter 5

And then you want to change something

The first eighty percent take ten minutes. The last twenty still take a day.

  • The AI knows WordPress. It does not know your website.
  • Framework, class names, nesting, style guide — all of it open questions.
  • Without rules it builds differently every time. And each result works on its own.
  • A craftsman who has lost his memory every morning.
  • One page always works. Ten sites over two years — who holds that together?

Speed moves the work. It does not remove it.

A real example, from this week

Alexander Schell had his agency's website built entirely by AI — design system, Figma library, components in Drupal — and wrote publicly about how it went. He is the best witness, because he is in favour.

  • What he reports

    Eight times faster than comparable projects, four times cheaper, because one person covered frontend, backend, UX and testing. That is not a small number, and I believe him.

  • Where he was needed

    By his own account not while building, but while approving and correcting. His examples: "No, the colour comes from the token." "No, that is the third card, the others go." "Compare it to the Figma layout, the spacing is off."

  • What is in the delivered markup

    The home page carries 792 values written straight into the classes, 129 distinct: twelve font sizes from 10 to 52 pixels and twelve colour values, typed by hand instead of taken from a token. The headings look healthy until you read the order: an h2 before the h1, then it jumps to h4. And the gold on the dark blue reaches 3.34:1 while being set at 10 to 13 pixels in 47 places, where 4.5:1 is required. Counted on 8 September 2026; he says himself the site is live and not yet finished.

  • Why this is not an accusation

    It is the confirmation. Exactly this work — deciding that one colour applies and the other three do not — stays with a human. It does not shrink when building gets faster. It becomes visible.

This is not a bug in the prompt. That is the nature of things: AI produces plausible variants. A design system needs the opposite — one decision that applies everywhere.