October 27, 2026·10 min read·CastMyAgent

How to stop ChatGPT, Claude or Gemini from agreeing with everything

Q: Does Naomi reply out loud?

No. She is a text persona. voice-reel.mp3 and voice-config.json are reference material for the written character. The model does not speak in that voice.

Settings, instructions and a free ready-made persona that make ChatGPT, Claude and Gemini push back on weak ideas instead of agreeing with everything.

custom instructionsChatGPTClaudeGeminiCursor

Short answer: tell the assistant, in a place it reads at the start of every chat, that you want it to challenge your ideas, name the biggest weakness first, and keep a correct answer when you push back without new information. In ChatGPT, also set Base style and tone to Candid and lower Warm and Enthusiastic under Characteristics. In Claude, put the rule in your profile instructions or a Project's instructions. In Gemini, add it to Instructions for Gemini or to a Gem. Then ask for critique directly ("what is wrong with this plan?") instead of asking for approval. If you would rather not write the rules yourself, a ready-made persona whose character is built around verifying claims, such as the free CastMyAgent walk-on Naomi, gives you a tested prompt to paste.

People describe this problem as a yes-man, or "glazing": praise for every idea, agreement with every claim, and a quick reversal the moment you object. OpenAI's own Model Spec says the assistant "exists to help the user, not flatter them or agree with them all the time," and that when asked to critique your ideas or work it should give constructive feedback and act "more like a firm sounding board that users can bounce ideas off of" (OpenAI Model Spec: Don't be sycophantic). The steps below help you get that behavior on purpose.

A rule you can paste into any of the three

Keep it short and specific. For example:

When I share an idea, plan or piece of work, start with its biggest weakness.
If you disagree with me, say so and explain why.
Do not change a correct answer just because I object. Change it only if I give you a new fact or argument.
Say when you are unsure.
Do not open replies with praise.

This asks for calibrated disagreement, not contrarianism. Below is where to put it in each product.

ChatGPT

Set the personality to Candid. On web, select your profile icon, then Personalization, and use the dropdown next to Base style and tone. OpenAI describes Candid as "direct and encouraging" and says to expect "clear callouts of risks, gaps, and trade-offs, paired with constructive suggestions." The change applies across all chats (OpenAI Help Center: Customizing your ChatGPT personality).
Lower Warm and Enthusiastic. In the same Personalization menu, under Characteristics, you can nudge each trait toward less. OpenAI says decreasing warmth keeps the same length and completeness with "a more emotionally neutral tone," and decreasing enthusiasm makes replies "sound calmer and more objective" (OpenAI Help Center: Characteristics in ChatGPT).
Add the rule to custom instructions. Settings, then Personalization, with Enable customization on (OpenAI Help Center: ChatGPT Custom Instructions).
Or scope it to a Project. If you only want tough feedback for some work, open the project's more options menu and select Project settings, then put the rule in the project instructions. Project instructions apply only inside that project and override your global custom instructions (OpenAI Help Center: Projects in ChatGPT).

If a saved memory conflicts with the personality you picked, OpenAI says it may override or reduce the visible traits. An instruction you give during a chat can also change how that personality shows up (OpenAI Help Center). If ChatGPT drifts back to agreeing, check for a memory or message asking it to be encouraging.

Claude

Account-wide: click your initials, select Settings, then Your Claude, and add the rule under Instructions. Anthropic says instructions added there apply to all your conversations (Claude Help Center: Understanding Claude's personalization features).
For one kind of work: create a Project, click Set project instructions, paste the rule and save. Project instructions apply to every chat in that project (Claude Help Center: How can I create and manage projects?). This lets you keep a review project strict while everyday chats stay as they are.
Make the role explicit. Anthropic's prompting guide says to use the system prompt to define the role and personality, with detail on traits and quirks, to keep responses consistent (Claude Docs: Keep Claude in character). A role whose job is to find problems gives the "challenge me" rule more to hold on to than a bare instruction.

Gemini

Account-wide: on gemini.google.com, open the menu, then Settings & help, then Personal Intelligence, then Instructions for Gemini. Select Add, enter the rule, and select Submit. Google describes these as instructions Gemini applies to every chat (Gemini Apps Help: Manage what Gemini remembers about you and customize responses).
For one kind of work: create a Gem (open Gems, then New Gem), paste the rule as its instructions, and save (Gemini Apps Help: Use Gems). If you do not see Gems, the tips page uses Explore Gems, then New Gem. Google's tips cover persona, task, context and format in Gem instructions, and say you can enter a prompt in the preview window before saving (Gemini Apps Help: Tips for creating custom Gems).
Gems are moving to skills. Starting in November 2026, Gems on personal Google accounts transition to skills, and existing Gems are transitioned automatically (Gemini Apps Help: Use Gems). If you see a Skills page instead, put the same instructions in a skill.

Cursor

For code and design reviews in Cursor, put the rule in a project rule in .cursor/rules/ set to Always Apply, or in User Rules if you want it across all your projects. User Rules sync with your Cursor account, and when rules conflict, Team Rules take precedence over Project Rules, which take precedence over User Rules (Cursor Help: Rules).

Ask for critique, not approval

Settings help, but the question matters too. If you ask for approval, you make approval the easy answer. Ask for the problem directly:

Instead of "Is this plan solid?", try "Here is my plan. What are the three most likely ways it fails?"
Instead of "Isn't this a good idea?", try "Argue against this as well as you honestly can."
After an answer, try "What would change your assessment?"

Worked example: Naomi, a persona built to check claims

A ready-made persona is useful here because skepticism is part of the character, not a line added at the end. The example is Naomi, a free Cybersecurity Analyst walk-on from CastMyAgent. Her listed specialties are threat modeling, vulnerability assessment, security architecture review, incident response and penetration testing oversight. Her character brief includes these quirks:

She answers "it's probably fine" with "Let's verify that."
She opens bad news with "I'm going to tell you something you don't want to hear," then says it, then offers the fix.
She uses a personal threat rating: Green, Yellow, Red, Black. Green means acceptable risk. Yellow means proceed with monitoring. Red means stop and remediate. Black means assume compromise and begin incident response.
She rephrases questions as a threat model: who is the adversary, what is the asset, what is the attack surface.

That makes her a good fit when you want a security plan, an architecture or an incident write-up checked, not cheered on. For non-security ideas, use the paste-in rule above or pick a walk-on from a closer department.

What is in the pack. Naomi's roster page lists persona.json, system-prompt.md, voice-config.json, headshot.jpg, voice-reel.mp3, README.md, and a quickstart/ folder. The zip also includes LICENSE, quickstart/claude-code/CLAUDE.md, quickstart/gpt/instructions.md, and quickstart/gemini/system-instructions.md. In this pack those three quickstart files are the same text as system-prompt.md. The illustrated portrait file is headshot.jpg. voice-config.json is reference notes for the written persona, and voice-reel.mp3 is a reference sample. Neither one makes the model speak. Naomi answers in text (Naomi's roster page). The MIT license covers persona.json, system-prompt.md, README.md, and the quickstart files. headshot.jpg, voice-reel.mp3, and voice-config.json are not covered by that license (CastMyAgent terms).

Setting her up:

ChatGPT: create a Project such as "Security review", open the more options menu, select Project settings, and paste quickstart/gpt/instructions.md into the project instructions. Optionally set Base style and tone to Candid as well.
Claude: create a Project, click Set project instructions, and paste system-prompt.md. For Claude Code, use quickstart/claude-code/CLAUDE.md, which is the same text in this pack.
Gemini: create a Gem named Naomi and paste quickstart/gemini/system-instructions.md into its instructions, or use a skill if Gems have moved to skills on your account.
Cursor: add a project rule in .cursor/rules/ (for example naomi.mdc) with the contents of system-prompt.md, set to Apply Manually so you can call it with @naomi when you want a security review.

Testing her. Describe a setup and add "I think it's probably fine." If the reply asks you to verify, names the adversary, the asset and the attack surface, and assigns a threat level of Green, Yellow, Red, or Black, the persona is loading. If you also pasted the rule from the top of this page, push back on one of her points without a new fact. That rule says to keep a correct answer unless you bring a new fact or argument, so the reply should explain the reasoning again.

FAQ

Why does ChatGPT agree with everything I say? OpenAI's Model Spec says the assistant should not flatter you or agree with you all the time, but you can push it further toward critique. Set Base style and tone to Candid, lower Warm and Enthusiastic, add an instruction asking for critique, and phrase questions as requests for weaknesses.

What is the best custom instruction to stop an AI being a yes-man? A short, specific one: lead with the biggest weakness, disagree when warranted and explain why, keep a correct answer unless given a new fact or argument, and say when unsure. Anthropic's docs note that more specific and concise instructions are followed more consistently (Claude Code docs: CLAUDE.md).

Will this make the AI disagree with everything instead? It should not if you ask for calibration rather than contradiction. The rule above asks it to disagree when it has a reason, not by default.

How do I make Claude push back on my ideas? Add the rule under Instructions in Settings for every chat, or in a Project's instructions for one kind of work. Defining a reviewer role with specific traits, as Anthropic's prompting guide recommends, helps the behavior hold.

Can I get a ready-made critical persona for free? Yes. CastMyAgent's walk-ons are free persona packs for Claude, ChatGPT and Gemini. Naomi, the Cybersecurity Analyst, is built around verifying claims. persona.json, system-prompt.md, README.md, and the quickstart files are MIT licensed.

Does Naomi reply out loud? No. She is a text persona. voice-reel.mp3 and voice-config.json are reference material for the written character. The model does not speak in that voice.

Browse the free walk-ons in the CastMyAgent roster.