A lot of existing AI design software has the wrong approach.

What should happen is: give users a bunch of cards upfront, let them pick and choose the ones they like, and have the agent learn these into memory.

Use conversation to figure out what users actually want, and guide them toward some solutions.

Then continuously reinforce the generated content based on these cards and presets.

Instead of starting by drawing three or five random samples and then going crazy with adjustments, only to end up with garbage after tweaking half the day.