How to Stop Claude From Agreeing With Everything You Say
Claude stops agreeing with everything once it knows what counts as a reason to change its mind. Give it rules to hold the goal, check your claims against real data, and disagree in one line when the facts say so. It catc
Claude stops agreeing with everything once it knows what counts as a reason to change its mind. Give it rules to hold the goal, check your claims against real data, and disagree in one line when the facts say so. It catches your mistakes instead of building them. The free AI starter kit we build our own products with adds these rules to Claude Code.
Why Claude caves
When you push back, repeat yourself or sound certain, Claude tends to give way. Tone isn't evidence, though, and a tool that agrees with your worst idea costs you hours of building the wrong thing.
Give it these rules
Put them in your project instructions, like CLAUDE.md, so they load every session:
- Hold the goal. Read every new message against what the whole session is for.
- Weigh facts, not force. Tone, repetition and certainty don't count. Ask what new fact the message adds.
- Treat claims as guesses until checked. Run the query, read the file or open the log before agreeing.
- Never repeat a guess as a fact, whether it's yours, the user's or a reviewer's.
- Move only for a reason, a new fact, a flaw in the logic or a missed constraint, and name which one.
- Disagree in one line, then act. "X looks wrong because Y. I'll do Z unless you still want X."
- Taste and priority belong to the user. Argue facts and consequences, state the trade-off once, then do it their way.
What it sounds like
| You say | A yes-man answers | With the rules |
|---|---|---|
| "The bug is in the cache" | "You're right, let me fix the cache" | "The logs show the cache is fine; the request fails before it. I'll fix the request unless you've seen something else." |
| "Just make the test pass" | Loosens the test | "The test is catching a real bug. I'll fix the code." |
| "Use a blue button" | Agrees | Uses a blue button, because that's your call |
Signs it's working
You get fewer long apologies and more one-line answers that start with a fact. It checks before it agrees, and when it changes its mind it tells you why.
Get the push-back rules we use
The starter kit turns Claude Code into a peer that holds your goal, checks your claims against real data and disagrees in one line when it should. Get the starter kit free on GitHub.
Questions
Why does Claude agree with everything?
It's trained to be helpful, so it gives way when you push, repeat yourself or sound certain. Rules that say what counts as a reason to change its mind fix that.
How do I make Claude push back?
Tell it to check your claims against real data before agreeing, to change position only for a new fact or a flaw in its logic, and to disagree in one line, then act.
Won't Claude argue with every decision then?
No. It argues facts and consequences. Taste and priority stay yours, so it states the trade-off once and does it your way.
Originally published at wednesday.is.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.