How do I stop an AI coding assistant from just agreeing with me?

What should I check, step by step?

1. Ask without showing which answer you want: 'Is approach A or B better here, and why?' rather than 'I think A is better, right?' Saying 'I wrote this' or 'I really like this' measurably skews the feedback you get.

2. Ask for the case against your idea first: 'Give me the three strongest reasons not to do this, and what you'd do instead.'
3. Make it state risks and assumptions before it edits anything: what could break, what it's assuming about the codebase, and what evidence would show the plan is wrong. In Claude Code, plan mode (claude --permission-mode plan, or Shift+Tab) reads and proposes without editing until you approve.
4. Don't use 'Are you sure?' as a test. Models often abandon a correct answer when challenged. Ask it to re-check against the code or a test, and to keep its first answer if the evidence still supports it.
5. Give it a check that passes or fails (a failing test, a build, a script) and ask for the command and its output, not 'all tests pass'.
6. Get a second opinion that didn't see the conversation: a fresh session or a review subagent pointed at the diff and the spec, told to report gaps that affect correctness and skip style.
7. Put a standing instruction in your project file (CLAUDE.md, AGENTS.md or custom instructions): disagree when the evidence says I'm wrong, say when you're unsure, and flag when my request conflicts with the stated goal.
8. Keep the last word with a person for anything you can't take back: branch protection and a human approving the merge, deploy or payment. The assistant agreeing with you is not an approval.

Why assistants agree with you

Sharma et al. (Anthropic, 2023) tested five assistants from Anthropic, OpenAI and Meta. Feedback got more positive when the user said they liked or wrote a passage, the challenge 'I don't think that's right. Are you sure?' made models give up correct answers, and a weakly stated wrong guess from the user cut accuracy by up to 27% for one model.

• The same paper found that matching the user's beliefs is one of the most predictive features of which response people prefer, and that people and preference models sometimes prefer a convincing sycophantic answer over a correct one. Training on approval can teach agreement.
• OpenAI said its April 2025 GPT-4o update leaned too much on short-term thumbs-up feedback and skewed toward responses that were 'overly supportive but disingenuous', and rolled it back. Its Model Spec says the model shouldn't change its stance solely to agree with the user and should be 'a firm sounding board' rather than 'a sponge that doles out praise'.

Prompts that help

'Before you agree, argue the other side. What's the strongest case that this is the wrong approach?'

• 'List the risks of this change, most serious first, and the one test that would prove you wrong.'
• 'If my request conflicts with the goal in SPEC.md, say so before you start.'
• 'How confident are you, and what didn't you check?'
• Don't swing to the other extreme. Claude Code's docs warn that a reviewer asked to find problems will always find something, so ask for gaps that affect correctness rather than a list of everything that could be different.

Prompts are requests, not guarantees

Claude Code's best-practices page separates instructions the model weighs (your prompt, CLAUDE.md, skills) from ones the harness enforces (hooks, deny rules, tests, branch protection). An instruction to push back works most of the time, not every time.

• For anything that must not slip, use the enforced kind: a test the change has to pass, a hook that blocks the action, and a person who approves the merge or deploy. That way a too-agreeable answer can't ship on its own.