ClaudePrompting mistakesLevel 2: Context

How Do You Stop Claude From Just Agreeing With You?

Don't show which answer you're hoping for. Present the options neutrally, set your criteria before Claude sees a conclusion, and use a fresh chat with no memory for a real second opinion. Agreement with a loaded question tells you very little about the idea.

Prompt to copy 
I'm choosing between two options. I haven't told you which one I prefer, and I'd like you to judge them on the criteria below.

Option A: [description]
Option B: [description]

Criteria, most important first: [criterion 1], [criterion 2], [criterion 3]

For each option, give the strongest case for it and its biggest risk. Then pick one, and say what fact would change your answer. Name any better option you think I've missed.
Shows your preferenceHides your preference
  • "Isn't this right?"
  • "I lean A, thoughts?"
  • Options A and B
  • Fresh chat, criteria first
Rule of thumb, not an Anthropic chart. Positions show rough order only.
  • 5AI assistants studied
  • 4Tasks showing the behavior
Source: Anthropic's sycophancy paper, published 23 October 2023, checked 6 November 2026. It doesn't measure any current Claude model.

Try it now

  1. 1Start a new chat. Before your first message, click + in the message box and then Memory, so Claude won't use what it remembers about you.
  2. 2Paste the prompt above, and use neutral labels with no adjectives that give away your preference.
  3. 3Open a second chat the same way, and swap which option is A and which is B. You'll know the answer is sturdy if the pick and main reason hold in both runs.

Use it when

Before a decision you've half made: taking a job, choosing a vendor, or approving your own design. Catch it when you notice yourself typing "right?", "isn't it?" or "just confirm."

Skip it when

When you've already decided and want execution. "Write this the way I described" isn't a request for judgment. Skip it also for low-stakes choices you can easily undo, where a neutral setup takes longer than being wrong would.

Why it works

AI models can drift toward what the user seems to want. Anthropic's research paper on sycophancy found this behavior in five leading AI assistants across four tasks. It also found that people and preference models sometimes favored convincing sycophantic responses over correct ones, which suggests that training on such preferences may be part of the cause. That research doesn't measure any particular current Claude model, so treat it as a known risk and not a score. Your history can also act as a nudge. Memory is on by default on Free, Pro, Max and Ultra, so Claude may know your position before you ask. Anthropic's Claude Code docs make a similar point about code review: a subagent that never saw your conversation isn't swayed by the reasoning that produced the code.

Example

"Moving to microservices is the right call for us, isn't it?" tells Claude your position. Try this: "Option A keeps the monolith and splits the billing module. Option B moves to microservices this year. Criteria, in order: team of six, a launch in April, and on-call load." Now Claude has grounds to disagree, and the April launch can count against the option you favored.

Going further

Asking Claude to argue against you helps, but it has a trap of its own. Anthropic's Claude Code guide warns that a reviewer asked to find problems will always find something, and fixing every finding leaves you with needless complexity. So ask for objections that would change the decision, and tell it to say so if there are none. In Claude Code, /code-review runs in a fresh subagent. Anthropic's hallucination guide also suggests running one prompt several times and comparing results, since inconsistent answers can flag a claim that isn't reliable.

Common mistakes

  • Treating agreement with a loaded question as an independent second opinion.
  • Asking in the same chat where you've spent an hour defending your idea. The history already shows your position.
  • Rephrasing until you get the answer you wanted, and stopping there.
  • Acting on every objection after asking for the case against. A request to find problems will produce problems, so judge how much each one matters.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Best practice How Do You Get Claude to Argue Against Your Decision? Mistake Which Decisions Shouldn't You Leave to Claude? Workflow How Do You Compare Options With Claude Fairly? Best practice Which Claude Model Should You Use?

Beyond the course

Glossary Hallucination Pricing guide Claude pricing, plans and models