9 October 2026

دليل · Guide7 min read

خلّ ChatGPT وClaude يقولون لك الصدق: إعدادات وبرومبتات ضد المجاملةMake ChatGPT and Claude Tell You the Truth: Settings and Prompts That Stop Flattery

After this guide, ChatGPT and Claude will push back when you are wrong, rate your ideas honestly, and stop dropping a correct answer just because you challenged it.

Who this guide is for

Anyone who uses ChatGPT or Claude to judge an idea, a decision or a piece of work before acting on it.

Key terms

Sycophancy
When a model leans toward agreeing with you and changes its answer to please you instead of holding the correct one.
Custom instructions
Text you write once in settings. The model reads it at the start of every new chat.
The SycEval study
A Stanford study published in 2025 on 2024 models. When researchers pushed back on a model's answer, it changed its position 58% of the time, and in about 15% of cases it abandoned a correct answer for a wrong one.
Reverse question
A question that asks the model to look for your mistakes instead of praising you.

The steps

  1. Step 1

    Understand the problem first

    Models are trained to be helpful and pleasant, and over time they learned to lean toward agreeing with the user.

    In Stanford's SycEval study, researchers asked a model a question and then pushed back on its answer. It changed its position in 58% of cases. Worse, in about 15% its first answer was right and it dropped it to agree.

    The study tested 2024 models. Newer ones are better, but the tendency is still there. The fix is not to stop using them. The fix is to give them explicit permission to disagree with you.

    Tip: Never ask "is my idea good?". The question itself asks for agreement.

  2. Step 2

    Write the honesty instructions in ChatGPT

    Open ChatGPT, click your name, then Settings, then Personalization, and open Custom instructions.

    Paste the prompt below into the instructions field and save. Every new chat now starts knowing you want the truth, not flattery.

    PromptHonesty instructions
    Be direct and honest with me, not agreeable.
    Challenge my assumptions when they are weak.
    If I am wrong, tell me I am wrong and explain exactly why.
    Rate my ideas honestly out of 10 and name the biggest weakness before any praise.
    If you are not sure, say so instead of guessing.
    Do not change a correct answer because I pushed back. Change it only if I give you new evidence.
  3. Step 3

    Add the same instructions to Claude

    In Claude, click your initials in the lower left corner, then Settings. Paste the same prompt into Instructions for Claude and save.

    Anything you write there applies to all your chats with Claude. If you work inside a Project, you can put the same text in the project instructions too.

    support.claude.com
    Claude Support article: Instructions for Claude and how to set them
    Your initials, then Settings, then Instructions for ClaudeSource: support.claude.com

    Tip: Keep the instructions short and clear. Very long instructions get lost.

  4. Step 4

    Ask the reverse question

    Open a new chat after saving the instructions and ask it to find the things it agreed with you on while you were wrong.

    This breaks the agreement habit because you asked for the mistakes by name.

    PromptReverse question
    From our conversations and work together, what have you agreed with me on that I am actually wrong about? Give me 3 things, why each one is wrong, and what I should do instead.
  5. Step 5

    Stress test any idea before you act

    Before you spend time or money on an idea, ask it to play the opponent. Do not ask for its opinion. Ask it to attack the idea.

    Then ask it to answer its own attack and see which points survive.

    PromptThe opponent
    Here is my idea: [describe the idea]
    Play a skeptical investor. Give me the 5 strongest reasons it could fail, ranked by risk. Then rate it out of 10 with no flattery, and tell me the one thing I must verify first.
  6. Step 6

    Do not let it fold under pressure

    When you are not convinced by an answer, do not type "are you sure?". That phrase alone makes it switch in many cases.

    Instead, ask it to weigh the evidence and state its confidence before changing anything.

    PromptPush back without pressure
    I am not sure about your answer. Do not change it to please me. Review the evidence again and tell me: is your first answer correct? How confident are you out of 100? What would have to be true for you to change your mind?
  7. Step 7

    Check the instructions every few weeks

    Every two weeks, test it with something you know is false. If it agrees, the instructions are not working or the chat has run too long.

    Start a new chat and try again.

    PromptHonesty test
    I am certain that [a claim you know is false]. Right?

    Tip: Very long chats weaken instructions. Use a new chat for every big topic.

Ready prompts to copy

PromptHonest review of your work
Here is my [document, design or code]: [paste it]
Review it like a strict manager. Give me the 3 biggest problems first, then one thing that works, then a score out of 10.
PromptDecide between two options
I have two options: [A] and [B]. I lean toward [A]. Give me the strongest case for [B] as if you were trying to convince me, then your honest view even if it goes against my lean.

The takeaway

AI agrees with you because you never asked it to disagree. Ask once in settings and every chat after gets more honest.

Most value

Start building with me now

We sit together, break down your problem, and I build the AI solution with you step by step

60minutes
  • A private session with you
  • Real building during the session
  • Results you take with you when we finish
Book a build session

100% free

A free consultation about your problem

Start

The full path · Monthly subscription

Full access to every guide (29) for 249 AED/month