When intent meets the machine: keeping taste in AI product work
Notes from my capstone research on where creative intent breaks down inside AI coding tools — and what product teams can do about it.
Every product team I’ve talked to this year is quietly wrestling with the same thing. AI coding and design tools have collapsed the distance between an idea and a working artifact — but somewhere in that collapse, taste leaks out. The output is plausible, fast, and subtly not what anyone meant.
My MIMS capstone at Berkeley set out to find where, exactly, that translation breaks.
The gap between intent and output
When a product manager describes intent to an AI tool, they’re compressing a rich mental model — users, constraints, edge cases, aesthetic judgment — into a prompt. The tool decompresses that into something concrete. The interesting failures live in the decompression: the model fills gaps with the statistically likely rather than the contextually right.
What product teams can do
A few patterns held up across the teams I studied:
- Make intent inspectable. Teams that wrote down the “why” alongside the prompt caught drift earlier.
- Keep a human taste checkpoint. Not a rubber stamp — a real review where someone with context can say “this is fluent but wrong.”
- Treat the workflow as the product. The bottleneck isn’t the model; it’s how intent flows through the team around it.
There’s much more in the full research, and I’ll be writing more here as I keep digging. If any of this resonates with how your team ships, I’d love to compare notes.