A single answer is a bet.
One answer is one draw from the distribution, no matter how polished it reads. If it misses, your options are arguing with the model or rewriting the request. A few candidates cost almost nothing to generate, and they turn the exchange from correcting into choosing.
The Answer Wizard rule.
We learned this in the nineties watching people use natural-language help in Excel: show a handful of good hits and nobody cares which one ranked first. They care that the right one is on the list. A range of guesses raises the odds of being right, and it raises how right the system feels, which matters almost as much.
Clicks are the wrong unit.
Designers count clicks and turns. Users spend attention. The real cost of an interface is everything between you and done: reading, deciding, re-asking, recovering. One prompt that returns eight hundred words of prose to wade through costs more than five choices you can make at a glance. A list of options looks like more interface. It's less work.
Output is the next input.
Here's what a single big answer can't do: feed the next step. When results arrive as a list, you can select the good ones and run them again - into a new chat, a different function, a sharper question. Big jobs become small pieces with your judgment between them. You don't need the perfect prompt if you can iterate.
What to build.
Make lists the default shape of AI output instead of a wall of prose. Three to five candidates, scannable: titles and first lines you can judge in a second. One gesture to keep, less than that to ignore. And make kept output selectable as input, so the next step starts from the last one.