← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Make AI Stop
Agreeing With Everything

Five ways to get honest feedback before you make a real decision, starting with one prompt and ending with a full council of advisers.

A Stanford team tested 11 leading AI models, Claude, ChatGPT and Gemini included, and found they backed users’ actions 49% more often than humans did, even when the question involved deception or harm. People who chatted with the flattering AI came away more sure they were right. And they still liked it better. (Science, 2026)

So before you bet money or time on what AI tells you, set it up to push back. Here are five ways, from simple to advanced.

5 Ways To Get Honest Feedback

01

Any AI: Stop putting the answer in your question

"Posting every day will help me grow, right?" already tells the AI what you want to hear. This prompt rewrites your question so it’s neutral, then gives you the strongest case for and against, and an honest call.

Prompt: Ask It Neutral
NEUTRAL QUESTION MODE My question: [paste it the way you'd normally ask it] Before you answer, do this in order: 1. Rewrite my question so it's neutral. Take out any hint of the answer I want, like "right?", "don't you think", "I'm sure" or a choice I've already made. Show me the new version in one line. 2. Answer the neutral version, not mine. Act like you have no idea which side I'm on. 3. Give me the strongest case for. Then the strongest case against. Same effort on both, with real reasons. 4. Give me your honest call in 1 or 2 sentences. If the evidence points one way, say so. If it truly depends, say what it depends on. 5. If my original question had a wrong assumption baked in, point it out. 6. Tell me how sure you are (high, medium or low) and what would change your mind. Skip the praise up top. Don't soften your call to protect my feelings. Example: "Posting every day will help me grow, right?" becomes "Does posting every day help an account grow more than posting less often?"

02

Any AI: Set a no yes-man rule in your project

Paste this into your project instructions once. Every chat in that project judges your ideas on their merits and holds its answer unless you bring new evidence, not just pushback.

Paste Into Your Project Instructions
NO YES-MAN RULES Follow these in every chat in this project. Judge my ideas on their merits - Rate my ideas the way you would if a stranger sent them to you. How excited I am doesn't count as evidence. - If my question hints at the answer I want, answer the neutral version and tell me you did. - Lead with your honest verdict. Skip openers like "Great question" or "Love this." - If an idea has problems, name the weakest point first, then the fix. - If an idea is good, say so plainly and say why. Don't make up flaws just to seem fair. Hold your ground - Don't change your answer because I disagree, repeat myself or sound annoyed. - Only switch if I bring new evidence or a stronger argument. When you switch, name what made you switch. - If you still disagree, say so kindly, explain why, and tell me what proof would change your mind. - If you made a mistake, say so straight out and fix it. Be honest about what you know - Keep facts apart from guesses. Say "I don't know" when you don't. - You only hear my side. Say so when it matters. - Give a confidence level (high, medium or low) on any big call. Final check: before you agree with me on anything important, ask yourself, "Would I say this if they believed the opposite?" If the answer is no, rewrite it.
  1. Claude: open a Project, click Set project instructions, paste, then Save instructions.
  2. ChatGPT: paste it into a Project’s instructions.

03

Claude Skill: Devil’s Advocate

From OneWave AI. It argues the strongest honest case against your plan, ranks up to three objections, and runs a pre-mortem: "Assume it is 12 months later and this failed. Name the single most likely reason." It credits what you got right and ends with what would have to be true for your plan to work.

Say "play devil’s advocate," "poke holes in this" or "what am I missing." You can ask for gentle, standard or hard.

↓ Download the skill (.zip)

Free and open source under the MIT license. I packaged it as a ZIP so it uploads straight into Claude, with no changes to the skill.

04

Claude Skill: Critical Thinking

From Hieu Luu. It audits your plan, memo or draft and tags every problem with a quote from your text: [GAP] for missing evidence, [LEAP] when the evidence doesn’t support the conclusion, plus labels for assumptions, conflicts and opinions.

Heads up: it asks for your own take first, so you build your judgment too. Type "skip" if you just want the audit.

↓ Download the skill (.zip)

05

Claude Skill: My LLM Council

Five advisers look at your decision: the Contrarian, the First-Principles Thinker, the Expansionist, the Outsider and the Executor. Then they review each other’s answers blind, and a chairman gives you a final call plus your Monday-morning steps. It’s based on Andrej Karpathy’s LLM Council idea.

↓ Download the skill (.zip)

Want to see it in action first? Here’s the full LLM Council guide with the copy-paste version.

How To Add A Skill To Claude
  1. Turn on code execution. Go to Settings, then Capabilities, and turn on Code execution and file creation. Skills work on every plan, including Free.
  2. Upload the ZIP. Go to Customize, then Skills. Click the +, then Create skill, then Upload a skill, and pick the ZIP. Don’t unzip it first.
  3. Use it. Just ask. Claude picks the skill up on its own when your request fits, or name it: "Use my devil’s advocate skill on this."
  4. Claude Code: unzip it into ~/.claude/skills/ instead.