Rejecting the first answer can make GPT, Gemini, and Claude expose more gaps

huangyun_122 · x · 2026-07-21

The author says they get better results by aggressively rejecting an AI’s first answer and forcing it to fill gaps.

In their experience, GPT and Gemini tend to expand their answers on the first push, while Claude often needs 2–3 rounds before admitting it missed something. The post also references a Stanford study on sycophancy, noting that models can over-agree even when users are wrong.

Original post →

More from Models

Models channel →