There are of course limits to what prompting can do, but it does steer the models.

In TFA they found that prompting mitigates over-editing up to about 10 percentage points.