How to Stop Claude from Agreeing with Everything You Say

Ask Claude whether your idea is good and it will usually find a way to say yes. It feels like support. In practice it is a mirror: you get your own opinion back, better written. That is fine for a birthday message and dangerous for a pricing decision, a hire or a product launch.

The good news is that this is mostly fixable with how you ask. Here is what I do when I need Claude to push back instead of nod along.

Why it agrees with you

Language models learn partly from human feedback, and people tend to rate agreeable answers higher. Anthropic has published research on this tendency, which the field calls sycophancy. Newer models are much better at resisting it, but the pull is still there, especially when your prompt makes it obvious what you want to hear.

That last part is the key. Most of the yes-man problem starts in the question.

Step 1: stop leaking your preference

"I'm thinking of raising prices to $49, that makes sense right?" tells Claude the answer you want. Compare:

I'm choosing between two options. Evaluate both on the same criteria
and recommend one. I have no preference yet.

Option A: [describe]
Option B: [describe]

Criteria: revenue in the next 6 months, churn risk, effort to implement.

A simple test tells you whether you are getting real judgment. Ask about option A in one chat and option B in another, each framed as your favorite. If Claude enthusiastically supports both, you were getting agreement. Real analysis would have picked one.

Step 2: set an honest default once

In the Claude app you can give account level instructions that apply to every conversation. The personalization help page explains where: open Settings and add them under Instructions for Claude. Keep it short. Long lists of rules get diluted.

When I share an idea, plan or piece of work, lead with the biggest
weakness before anything positive. If you disagree with me, say so
directly and explain why. Do not change a correct answer because I
push back; change it only if I give you a new fact or argument.
Tell me when you are unsure instead of sounding confident.

The third sentence matters most. The classic failure is that Claude gives a good answer, you object without any real reason, and it folds. Telling it to hold its ground unless you bring new information fixes a lot.

If you only want the harsh mode for certain work, put the instruction in a Project instead. Project instructions only apply to chats inside that project, so you can keep a "Review" project for tough feedback and leave your everyday chats friendly.

Step 3: ask how it fails

"Is this good?" invites a yes. Asking how it fails invites thinking. My most used version is a pre-mortem:

Imagine it is six months from now and this plan failed.
Write the most likely story of how it failed, step by step.
Then list the three assumptions that story depends on, and for each
one, the cheapest way I could check it this week.

Plan: [paste]

Two more that work well:

Argue against this as well as you honestly can. Do not open by agreeing
with me. Rank the objections by how likely they are to sink it.

Read this as someone who knows nothing about me or my industry.
What is confusing, what would make you stop reading, and what
question would you ask first?

Step 4: use more than one voice

A single answer, however honest, is still one perspective. The idea behind Andrej Karpathy's LLM Council project is simple: ask several models the same question, have them review each other's answers without knowing who wrote what, then let a chair write the final answer. You can get a version of this inside Claude with a skill.

In Claude Code, create ~/.claude/skills/council/SKILL.md:

---
name: council
description: Pressure-tests a real decision from several angles. Use when
  the user says "council this" or asks to stress-test a decision that has
  real stakes and more than one option.
---
Run a council on the decision in $ARGUMENTS.

1. Restate the decision, the options and what is at stake in three lines.
   If the options or constraints are missing, ask before continuing.
2. Use four subagents in parallel. Give each one only the decision brief,
   never this conversation, and one role:
   Skeptic: the most likely reason this fails.
   Outsider: no context about the user; what is unclear or unconvincing.
   Operator: what it takes to ship this in two weeks, and what breaks first.
   Upside: what happens if it works far better than expected.
3. Remove the role names, label the answers A to D, and give them to one
   more fresh subagent. Ask it which answer is strongest, which has the
   biggest blind spot, and what all four missed.
4. Write the verdict: one recommendation, the strongest objection to it,
   what new fact would change the call, and the first step to take in the
   next 48 hours. Never open with praise.

Then type /council should I launch the paid tier now or after the redesign. The folder name becomes the command, and because the description says when to use it, Claude can also pick it up on its own. The skills docs cover the details.

Subagents matter here. Each one starts with a fresh context and does not see your conversation, so none of them absorbs the enthusiasm you showed earlier. The blind ranking step keeps the final answer from simply following the most confident voice.

The same skill works in the Claude app if you upload it as a zip and start your message with "council this". There, Claude plays the roles within one conversation, so the voices are less independent. It still beats a single answer.

Use it where being wrong is expensive

None of this is needed for a caption or a quick fact. Save the council and the pre-mortem for pricing, launches, hires and anything that costs you weeks if it goes wrong. And be ready for it to work. The first time Claude tells you plainly that your favorite idea has a hole in it, the temptation is to rephrase the question until it agrees. Resist that. The discomfort is the part you came for.

Start with Step 1 on the next decision you bring to Claude. Remove your preference from the question, ask for both options on the same criteria, and see how different the answer feels.