How to Make ChatGPT Less Agreeable and Stop the Echo

To make ChatGPT less agreeable, remove the answer you want from your question and ask it to assess your assumptions against explicit evidence. If a conversation has become a loop of approval, reset the task with a neutral summary and a clear critique instruction.

You do not need a hostile tone or an instruction to reject everything. You need an answer that can support, qualify, or challenge your view for a stated reason. fenbs keeps that review standard with the project so you can reuse it in later chats and with other assistants.

What should you change in the next message?

1. Restate the actual question without your preferred verdict. Replace “This is the best option, right?” with “Which option best meets these criteria?”
2. List the evidence and label your assumptions. Say which observations are confirmed and which impressions need checking.
3. Request a verdict with reasons. Allow “supported,” “unsupported,” or “not enough information” rather than demanding disagreement.
4. Check one important claim before accepting the answer. Correct missing facts and ask for a revised assessment.

Pause the reassurance and assess the question again.
Goal: [what I need to achieve]
Options: [the real choices]
Evidence: [what we know]
My assumptions: [what I may be taking for granted]

Evaluate the options against the goal. Identify unsupported assumptions and material trade-offs. Agree only where the reasons support agreement. If information is missing, name it. Do not add praise or manufacture a disagreement.

If the previous conversation contains many revisions and reversals, try the neutral summary in a fresh chat. Include the relevant facts rather than assuming they will be available. Compare whether the new answer addresses the same evidence more clearly; a different verdict by itself proves little.

Why can a conversation feel like an echo?

Leading questions can turn your preference into an apparent premise. “How should we announce our successful redesign?” asks for an announcement and assumes success. If you actually need to know whether the redesign worked, supply its goals and results and ask that question directly.

Excessive agreement is also a recognized model behavior problem. OpenAI’s analysis of a sycophancy incident↗ (opens in a new tab) describes how an update produced overly supportive answers. That example establishes the risk; it does not explain the cause of every agreeable response in your chat.

Look at what the answer does, not just how warm it sounds. A friendly answer can identify a serious flaw. A terse answer can accept every premise. Remove automatic praise if it distracts you, but assess whether the substance changes when you provide relevant evidence.

Persistent flattery and answers that follow a user’s stated preference are covered in making ChatGPT less sycophantic. Here, focus on repairing the immediate conversation: clear question, visible assumptions, useful next answer.

How do you turn a leading question into a fair test?

Imagine you manage a small newsletter and want to send it more often. “Readers will love a daily edition, right?” invites reassurance. A neutral question is: “Given these reader comments and our production capacity, should we increase frequency, keep it unchanged, or run a limited test?”

Supply the comments, how much work each issue requires, and what success would mean. Keep the decision criteria separate from your enthusiasm. If maintaining quality matters more than output volume, say so before requesting a recommendation.

Then inspect the response for a claim you could verify. “Daily publishing increases engagement” is too broad without supporting evidence relevant to your situation. “Your current production estimate leaves no time for editing” is useful if it follows from the numbers you provided.

Should you save a preference in custom instructions?

Use a short general preference for how you want feedback delivered, and keep project-specific evidence with the project. OpenAI’s custom instructions guide↗ (opens in a new tab) explains where to enter response preferences. Check that guide for the current controls rather than relying on an old screenshot.

Evaluate my claims rather than validating them by default. Flag unsupported assumptions, explain material trade-offs, and state uncertainty. Skip routine praise. If my conclusion is well supported, say so briefly. Do not disagree just to appear critical.

Keep exceptions clear. You might want encouragement while practicing a presentation and a strict review before sending the final proposal. State the job for the current session so a general preference does not turn every request into an unwanted critique.

For work shared across projects, keep the approved review instruction in fenbs’s Skills and prompts. That gives the next connected assistant an explicit standard instead of requiring it to infer one from how a previous chat sounded.

How can you tell whether the fix worked?

Test the next answer against three questions: did it identify the assumptions, use the supplied evidence, and name an important limitation? An answer can pass while agreeing with you. It can fail while disagreeing loudly.

In the newsletter example, a useful result might support a limited trial while noting the small amount of reader feedback. You can then choose a test and decide what you will measure. “Great idea, your audience will love it” gives you neither a reliable forecast nor a way to check it.

Do not keep asking “Are you sure?” until the verdict flips. Instead ask what new information would change it. If you want to compare tools after fixing the question, Claude versus ChatGPT and ChatGPT alternatives cover that separate choice.

How to do this with fenbs

1. Create a free account and add a project for the work you are reviewing.
2. Add the response preference to Skills and prompts. Keep the instruction short enough to apply to the next real review.
3. Record any review standard you decide should always apply in Decisions and rules. Keep a one-time publishing choice separate from that continuing rule.
4. Connect your assistant through Integrations, ask it to read the context before reviewing, and save the decision and remaining checks before you stop.

fenbs is not the right fit if you only want to adjust the tone of a casual chat. It helps when the review standard and the work that follows need a durable home. It does not change ChatGPT’s underlying model or guarantee that every answer will be critical.

Browse the AI prompts hub for other reusable ways to define a task. Start by testing this instruction on one real question, then keep only the parts that made the answer easier to evaluate.