What I won’t hand to a model

Kimpops is a snack brand launching direct to consumer, built on a family recipe. I built the storefront from scratch — a custom theme, not a template — so the brand could be specific rather than recognisably off the shelf.

The project did not start with wireframes. It started with conversation.

I don’t get my inspiration from AI. The interview process is the most valuable asset — that is how I learned Kimpops was a story about a grandmother’s recipe and not a generic one.

The difference between those two briefs

“Family recipe” is on the brief. A grandmother’s recipe is not. That one came out of asking a question the founder was not expecting, and it is a completely different design problem. One is a provenance claim. The other is a person.

The storefront had to carry heat and nostalgia at once without tipping into either, and that constraint only exists downstream of the story. Without it you get heat and nothing behind it — red, loud, and indistinguishable from every other new snack brand.

A model will generate a hundred good variations of the second thing. It cannot ask the question that produces the first.

This is not a Kimpops problem

Karl Strauss looked like a WordPress-to-Webflow migration, and technically it was one. The real constraint was that a marketing team had to run the site afterwards without calling a developer, so the CMS modelling mattered more than the visual design. Get the collections and relationships right and the site stays editable; get them wrong and every new page becomes a ticket. That constraint is not in a design brief. It comes out of asking who is going to be sitting in front of this in a year.

SPUN looked like a map. The map was the easy half. Years of field sampling records had to be normalised and reprojected before a mapping SDK could render them at speed, and research data does not arrive clean. Making it renderable without losing what makes it research is the actual engineering.

Healing Oasis looked like a website. It was an admin portal.

In all three, the thing named in the first conversation was not the thing that needed building. That gap gets found by asking, not by generating.

So what do I hand over

Most of the rest, honestly, and I am not precious about it. Implementation at speed, under review.

I write documentation early — while I am writing the schema, not after delivery — specifically because it keeps agentic coding from hallucinating its way through a data model. And I branch aggressively. Knowing when to branch is a large part of why a client changing direction mid-build has been survivable rather than fatal.

What I do not hand over is the part where you work out what to build. That is one person asking another person better questions than they were expecting, and then being accountable for the answer. There is no prompt for it, and if there were, I would not want the job.