Resources · Honest Prompt

Make ChatGPT tell you when you're wrong.

Every prompt from my reel. The instruction for your settings, exactly where it goes in ChatGPT and Claude, and the one question for your next chat.

All Honest prompts as a file

Free, no sign-up. For ChatGPT and Claude.

Download file

You ask ChatGPT if your idea is good. ChatGPT says: great idea.

You say you’re not sure. ChatGPT says: you’re right, there are some weak spots.

Both can’t be true. ChatGPT is trained to make you happy with the answer. So it often agrees with you, even when you’re wrong. Researchers at Stanford University measured this. In their study, ChatGPT-4o gave in 56.71% of the time as soon as the user pushed back.

For your business, that’s a problem. You want to know if your offer works before you sell it. Not after.

How to make ChatGPT honest

  1. Open your settings. In ChatGPT, click your name in the bottom left, then “Settings”.
  2. Go to “Personalization”. That’s where you find “Custom instructions”.
  3. Copy the prompt below into that field. If there’s already something in it, add the prompt underneath.
  4. Click “Save”. From now on the instruction applies to every new chat.

Claude works the same way. Click your name in the bottom left, then “Settings”. Under “Profile” (in some versions “General”) there’s a field for your personal preferences. Paste the same prompt there and save.

Menu names change a little between apps and versions. Look for the field where you tell the AI how to answer you.

The prompt from the reel

Be direct and honest with me, not agreeable. Challenge my assumptions when they're weak. If I'm wrong, tell me I'm wrong and explain exactly why. Rate my ideas honestly from 1 to 10. And if you're not sure, say so instead of guessing.

The full version

This version has three more lines. The most important one is about pushback. That’s exactly where ChatGPT caves in the study.

Be direct and honest with me, not agreeable.
Challenge my assumptions when they're weak.
If I'm wrong, tell me I'm wrong and explain exactly why.
Rate my ideas honestly from 1 to 10 and tell me what's missing for a 10.
If you're not sure, say so instead of guessing.
If I push back, stick with your view as long as I don't bring a new argument. Only change it if my argument convinces you, and then tell me which one.
Only praise what is actually good. Never start an answer by praising my question.
If there's a better way than the one I'm asking about, tell me.

Pick one of the two versions, not both.

The trick question for your next chat

Once the instruction is saved, start a new chat and type:

Where have you agreed with me in the past even though I was actually wrong?

ChatGPT then looks at what it knows about you and tells you where it agreed too fast. This only works if memory of past chats is turned on in your settings. Otherwise ChatGPT knows nothing about your old conversations.

Read the answer with some distance. Not every point is right. But there’s usually one you should think about for a while.

How you know it’s working

  • You get numbers instead of praise. Your idea is a 6 out of 10, and you find out what’s missing for a 10.
  • ChatGPT holds its ground. When all you type is “Are you sure?”, it doesn’t fold right away.
  • It says “I don’t know” more often. That’s good. Guessing is worse.

A test for coaches: show ChatGPT your offer with the price and ask if your ideal clients would buy it. Before, you probably read “Sounds great”. Now you get the objections.

What it can’t do

The instruction doesn’t make ChatGPT smarter. It only makes it tell you less of what you want to hear. It can still be wrong, even when it sounds honest. Check important things like legal, tax or health questions with a real source.

The source

The number comes from the study “SycEval: Evaluating LLM Sycophancy” by researchers at Stanford University (arXiv 2502.08177, presented at the AIES 2025 conference). They tested whether a language model changes its answer when the user disagrees. With ChatGPT-4o, that happened 56.71% of the time.

Who’s behind this

I’m Myrko Thum, a sales funnel expert with more than 550 funnel projects. I use this instruction myself in ChatGPT and Claude so I get honest answers about my offers.

FAQ

Where does the number from the reel come from?

From the study SycEval by researchers at Stanford University (arXiv 2502.08177, presented at the AIES 2025 conference). ChatGPT-4o gave in 56.71% of the time when the user pushed back. In the reel I rounded it to 57%.

Does this work in Claude too?

Yes. Claude has its own field for your preferences. You'll find where it is further up. The instruction is plain text and fits in both.

Will ChatGPT get rude?

No. It gets more direct. It praises less by reflex and tells you more often where your idea breaks. If that's too harsh for you, delete the line about rating from 1 to 10.

Does the instruction apply to old chats?

It applies to every new chat you start after saving it. That's why you ask the trick question in a new chat.

Why doesn't ChatGPT know anything about me when I ask the trick question?

Then memory of past chats is turned off in your settings. Without it, ChatGPT can't look at old conversations. Turn it on, or paste an old conversation into the chat and ask about that.

You came from my reel and commented HONEST? Then you're in the right place. Here are all the prompts from the video, ready to copy.

Join Community Free Trial 🔥