The second-opinion habit

Models are agreeable. Ask for a plan and you get a plan. Ask whether the plan is good and you'll often get encouragement. This is pleasant and it is the main way these tools quietly make people's thinking worse: the first plausible answer arrives fast, sounds composed, and stops the search.

The correction is a habit rather than a prompt. After any answer that matters, run a second pass whose only job is to attack the first.

Three moves that work

Invert the role. "You wrote that plan. Now argue, as the person most likely to be hurt by it, why it fails." The role change matters more than the instruction — asking "are there any problems?" invites a list of hedges, while asking it to occupy an opposed position produces actual objections.

Ask for the assumption, not the flaw. "What has to be true for this to work? Which of those is least supported?" Most bad plans aren't logically broken; they rest on a premise nobody stated. This surfaces the premise.

Force a ranked disagreement. "Give me the three strongest objections, ordered by how much they should change my mind." Ranking is the part that does the work. An unranked list of concerns is noise; the ordering is a judgment you can argue with.

Where it's wasted

Don't do this for tasks with a checkable right answer. Reformatting, extraction, translation, code that either compiles or doesn't — a critique pass adds nothing but a second opinion on something already verifiable.

Save it for decisions: what to build, what to prioritise, how to frame something, whether an argument holds. The pattern is simple — if you'd want a colleague to look at it before you committed, run the second pass.

The version I actually use

Now take the opposite position. What's the strongest case that this is wrong, and what would I need to observe to find out?

The last clause is the useful half. An objection you can't test is an opinion. An objection with an observable consequence is something you can go and check.