The creative team was waiting weeks for product imagery. I wrote the PRD, designed the tool, prototyped it myself, and shipped it with two engineers — the first thing I've owned from problem statement to production.
A jewellery catalogue turns over constantly, and the creative team was the constraint on everything downstream — campaign pages, ads, app banners, category art. A lifestyle image meant scheduling a shoot, which meant a two-to-three week lead time for something that would run for four days.
Generic AI image tools were already being used unofficially, and the output was unusable: the product came out wrong. Stones in the wrong cut, clasps invented, gold that didn't match the actual SKU. Nobody could ship it, so everyone went back to waiting.
Watching the creative team work, the pattern was obvious: they never wanted the AI to invent the jewellery. They wanted to keep the real product photograph exactly as shot and change the world around it — a model, a setting, a season, a mood.
So imaGenie isn't a prompt box. You start from a catalogue SKU, and the product pixels are treated as fixed. The generation happens around them. That single constraint is what made the output shippable, and it came from an hour of watching rather than a feature request.
The whole flow is designed so a creative who has never written a prompt gets a usable result on the first attempt, and a good one by the third.
I don't have an engineering background. What I have now is AI-assisted tooling, which was enough to build a working prototype wired to a real generation API — not a clickable mock, an actual thing that produced actual images and failed in actual ways.
That changed the conversation with engineering completely. Instead of debating whether a fidelity flag was feasible, we looked at my version doing it badly and talked about how to do it properly. Scope arguments got shorter because the disagreements were about behaviour we could both see.
It also changed what I caught. Generation latency made my first design wrong — a four-up grid that appears all at once is a very different experience when it takes eleven seconds. I only found that because I sat waiting for my own prototype.
The honest caveat: adoption is still uneven. Two of the creatives use it daily, one doesn't trust it yet, and I think that's a legitimate position — the fidelity flag is a heuristic, not a guarantee, and I haven't solved how to communicate that confidence properly.
What I'd repeat everywhere: prototyping in the real material. Every meaningful decision here came from using the thing, not from reviewing a frame of it.