My Own Pipeline Served Me Slop

The first full run of a new tool cost me nothing, and for a moment it looked like the output was worth exactly that.

The tool generates a preview website for a real small business, a mock of what their site could be. Before wiring in real credentials I ran the whole pipeline against mocked services, a dress rehearsal with no bills attached. The plumbing held. Leads came through, a mock generated, pages rendered end to end.

Then I actually looked at the preview. Generic. No design elements, nothing tied to the business’s category, nothing a stranger would recognise as being about them. My own product, and my first reaction was that it screamed AI slop.

Here is the uncomfortable part: the pipeline was working as built. I had given the generator freedom where it needed constraint. It was composing pages from the model’s idea of a website instead of drawing on the template library I had already designed, real layouts with a point of view, matched to the category of the business receiving them.

That preview is a pitch. To the person opening it, a generic page says nobody looked at you, the same way a letter opening with Dear Valued Customer says it. Mass-produced is legible to everyone, instantly.

So the fix was not a better prompt. It was grounding the generation in artefacts that carry my taste: the templates, the category styling, the design decisions made before the model got involved. The model supplies volume. The point of view still has to be mine.

Slop is the default. You have to build your way out of it.