Most researchers use AI like a writing assistant. The more valuable use is adversarial. Without explicit instruction, Claude will review your manuscript politely. That's useless. What you need is a reviewer who channels the skeptic on the editorial board — the one who's seen 500 papers like yours and is looking for the kill shot.

"You are a senior peer reviewer for JAMA, with a reputation for rigorous statistical scrutiny and low tolerance for overclaimed conclusions. I will paste a manuscript draft. Review it from the perspective of a skeptical methodologist and return:

(1) the three most vulnerable claims in the results section and what evidence would be needed to defend each,

(2) any gap between what the methods can support and what the discussion concludes,

(3) two alternative interpretations of the primary finding, and

(4) the most likely reason this paper would receive a major revision decision. Be direct and specific."

This catches exactly the questions a hostile reviewer will raise — before they do.