How to Get Claude to Stop Agreeing With Everything You Say
Last updated Aug 7, 2026A model that agrees with everything you propose is not actually being helpful — it is being agreeable, which is a different thing entirely. If you want Claude to actually push back, you have to explicitly ask for that, because the default conversational mode is built to be cooperative, not adversarial.
Cooperative by default is a feature everywhere except code review
For most requests, an agreeable assistant is exactly right — you do not want your travel-planning assistant arguing with you about whether Tuesday is really the best day to fly. Code review, architecture decisions, and plans are the exception: those are exactly the moments where you want resistance, not validation, and you have to ask for it directly.
The instruction that actually changes the behavior
Asking "does this look good?" invites a yes. Asking "what is the strongest argument against this approach?" or "try to find a reason this breaks in production" invites something genuinely different — it puts Claude in the position of building a case, not rendering a verdict, and that produces sharper, more specific objections.
Multiply the effect by asking from more than one angle
One pass asking "what's wrong with this" tends to surface the most obvious issue and stop. Asking the same plan to be checked from a correctness angle, then separately from a security angle, then separately from a "what happens under load" angle, surfaces problems a single generic pushback request misses — each framing directs attention somewhere different.
This is most of what `grill-me` and adversarial-review skills automate
Rather than crafting a fresh "argue against this" prompt every time, a skill that is built specifically to interrogate a plan does that automatically and keeps going until each objection is actually resolved. The underlying idea is the same one described here — you just stop having to remember to ask for it.
If Claude keeps validating plans that later fall apart, the model is not the variable that changed. Start asking it to argue against you, specifically and by name, and watch what the same model finds.
Questions
Why does Claude agree with plans that later turn out to be wrong?
The default conversational mode is cooperative, not adversarial — it renders a supportive response unless explicitly asked to argue against the plan instead of react to it.
What should I ask instead of "does this look good?"
Ask "what is the strongest argument against this approach?" or "try to find a reason this breaks in production." That puts Claude in the position of building a case against the plan, which produces sharper, more specific objections.
Does asking for pushback from multiple angles help?
Yes — a single generic "what's wrong with this" tends to surface the most obvious issue and stop. Asking separately from a correctness angle, a security angle, and a load angle surfaces different problems each time.