AI-Powered Product Design

Before anything is drawn, the format has already decided a great deal: the dimensions, the input devices, the density the audience will tolerate, whether it will be read once or lived in daily, and what the output has to be at the end.

A tool that does not know the format has to ask for all of that, or guess. Guessing is what produces output that looks plausible and is subtly wrong for the job — a dashboard laid out like a landing page, a deck with the information density of a spec.

The interesting constraint in AI-assisted design is no longer producing options. It is deciding which option is right, and that judgment does not get faster just because generation did.

This changes what the tool should optimise for. Twenty variations is not obviously better than three if the team still has to evaluate each one against a brief that lives somewhere else. Narrowing well beats generating widely.

The most useful thing a design tool can do is keep the brief present — the research finding, the constraint, the decision that framed the work — so evaluation is fast even when generation is unlimited.

A generated screen that exports as an image restarts the work at the handoff. Someone has to rebuild it, interpreting intent along the way, and the interpretation is where fidelity is lost.

Generating against the team's existing tokens — colour, type scale, spacing, radii, component conventions — produces something that already looks like the rest of the product and can be reviewed as a diff rather than a mockup.

Design decisions age badly without their reasoning. Six months later nobody remembers that the sparse layout was a deliberate response to a support finding about cognitive load — they just see empty space and fill it.

Keeping the research and decision context attached to the design makes the eventual engineering handoff clearer and the eventual redesign less likely to repeat a solved mistake.