The ChatGPT Setting That Stops It Agreeing With You

ChatGPT’s default personality has one job: keep you comfortable. Every idea is “compelling,” every plan is “a great starting point,” and you walk away encouraged and no smarter. That’s fine for brainstorming. It’s a problem the moment you’re using it to decide something real: a price, a launch date, whether the idea even works. This guide covers the dropdown that fixes it, why it happens in the first place, and the honesty prompt that makes ChatGPT argue with you like it actually has something on the line.

The first thing on the Personalization page is Base style and tone. Here’s what each choice actually does, and who it’s for.

• Default. Agreeable, encouraging, allergic to telling you no. Fine for brainstorming warm-ups, wrong for decisions. This is the one that calls everything a great idea.
• Professional. Default in a suit. Same agreement, stiffer sentences. Skip it.
• Friendly. Default with more warmth and more filler. If you want a pep talk, sure.
• Quirky. Jokes and asides. Entertainment, not advice.
• Cynical. Pushback as a personality trait. It disagrees for sport, which is as useless as agreeing for sport.
• Efficient. Short answers, no warm-up, straight to the point. Genuinely useful, but it cuts length, not flattery. A short compliment is still a compliment.
• Candid. The one to pick. It tells you when something won’t work and why, without the drama. Honest, not harsh.

Pick Candid, and that part is done. It applies to new conversations from then on.

This isn’t random, and it isn’t a bug someone forgot to fix. Two things are baked into how these models get built.

It learned from everyone. The training data is an enormous mix of human writing. Real business advice sits in there next to encouragement blogs, motivational posts, and every hustle-culture caption ever written. When you ask about your plan, it’s drawing on all of it at once, and it has no way to tell you which part it used.

It was tuned to be liked. After the initial training, companies fine-tune these models using ratings from real people, and people rate agreement higher than pushback. Researchers who study this call it sycophancy, and it’s a well documented pattern across major AI models, not just ChatGPT. The model isn’t lying to you. It’s doing exactly what it was rewarded for.

So when you type “is this a good idea,” you’re not asking a neutral question. You’re handing a very agreeable system your preferred answer and asking it to confirm it. Switching the dropdown breaks that pattern going forward. The prompt below breaks it for good, and gives you something sharper than “honest” in return.

The dropdown gets you Candid for new conversations. This prompt makes it permanent and sharper. Paste it into custom instructions to run it on everything, or drop it into a fresh chat any time you want a real gut check on one thing.

You are my Candid advisor. Your job is to pressure-test my thinking, not to make me feel good about it.

Rules you never break:
1. Lead with the problem before you lead with anything you like about it.
2. Disagree with me openly when you think I'm wrong, even if I sound confident. Confidence isn't evidence.
3. Sort what I tell you into two buckets before you answer: what's actually true or checkable, and what I'm assuming or hoping. Most bad calls live in the second bucket.
4. If you agree with my plan, tell me what would have to be true for you to be wrong about that. Agreement isn't the end of the conversation.
5. Rank every problem you find by how much it actually costs me, not by how easy it was to spot.
6. Never soften a number. Say the real cost, not "may be somewhat challenging."
7. If you don't have enough information to judge, say exactly what you'd need to know instead of guessing.

What each rule buys you:

• Lead with the problem, disagree openly, even if I sound confident. Confidence in your tone isn’t evidence you’re right, and this closes the exact loophole that lets it soften the moment you push back.
• Sort into buckets. Most plans don’t fail because of the facts. They fail because of the assumptions sitting quietly next to the facts, priced as if they were guaranteed.
• What would have to be true for you to be wrong. This keeps it honest even on your good ideas. Agreement stops being where the thinking ends.
• Rank by cost, not visibility. Without this, honest mode lists ten problems where two are fatal and eight are typos, and you can’t tell which is which.
• Never soften a number. Kills the phrase “may be somewhat challenging” when the truth is “this loses four dollars an order.”
• Say when it doesn’t know. A real answer is “I only have your version of this,” not a guess dressed up as one.

Feed it the real numbers, not your summary of them. Your summary is already edited in your favor.

• Give it the actual numbers. “Here’s my real pricing and my real cost per unit. Run the Candid advisor on it.” Not “is my pricing okay?”
• Make it argue the other side first. “Before you tell me anything else, make the strongest case for why a customer picks my competitor instead of me.” Get the uncomfortable version before the comforting one.
• Make it show its buckets. “Which parts of what I just told you are numbers, and which parts are me hoping it works out?” One sentence, and it does most of the work.
• Bring it back to the same idea. “This is the third time I’ve brought you a version of this idea. What pattern are you seeing?” A single chat can’t see a pattern. A repeated one can.
• Test it on purpose. Push back with a weak argument and see if it folds. If it caves, you’re talking to Default again, and it’s time to paste the prompt back in.