The framing matters more than the prompt. Ask an AI to argue against your idea and you'll get hedged objections wrapped in caveats - technically pushback, but soft enough that it doesn't actually challenge anything. The model is trying to be helpful without deflating the person whose idea it's critiquing.
A sharper approach: write two versions of the same thing yourself, even if the second is deliberately rough, then ask Claude to score both and explain which is weaker. Now the model isn't criticizing your work - it's judging a competition between two artifacts. That framing shift changes how the feedback lands.
The reason this works is rooted in how conversational AI handles perceived social risk. Asking "argue against this" puts the model in a position where pushing back hard might feel aggressive. Asking "which of these two is weaker and why" removes that pressure - it's an evaluation task, not a confrontation. The model doesn't need to worry about deflating you; it just needs to be accurate.
This works for copy, strategy documents, article structures, pitches, product descriptions, and code architecture decisions - anything where you can create or fabricate a second option. The worse your decoy version, the more obvious the contrast becomes, but that's not really the goal. You want enough tension between the two versions that the model has something real to say about the gap.
A few setups that produce particularly useful feedback: write your best email draft and a hedge-everything version, then ask which reads as more confident and why. Write your actual pitch angle and a generic version, then ask which makes a stronger case. Write your real product description and a bland competitor-style one, then ask which is more specific.
The broader principle holds beyond critique: AI performs better on evaluation tasks than on confrontation tasks. Judges don't soften verdicts the way conversation partners do. If you want an honest assessment, set up the conditions for an honest judgment.