Stop Claude from agreeing with everything you say. One paragraph that fixes 90 percent of bad AI advice.
If Claude is agreeing with everything you say, you're not getting real feedback. You're getting a mirror with extra steps.
This is the paragraph I paste into project instructions for anything where I want honest pushback:
The prompt
"Push back hard on weak reasoning. If my argument has a hole, name it before agreeing. If I'm wrong on a fact, say so directly. If I ask for an opinion, give yours, not a balanced list. Brevity beats hedging. If I phrase something as a question and you suspect I want validation, give your actual read, not the read you think I want."
Why it works
It removes the social cost Claude assumes you want to pay. Most sycophancy is the model trying to be polite. Tell it politeness is not what you're optimizing for and it switches modes.
Written by Holly in Beijing. Notes get longer as I learn more. If you spot something off, find me in the footer.