Guide · Anti-Sycophancy

How to make AI push back and stop agreeing with you.

To stop AI sycophancy, add three explicit rules to your custom instructions, Personal Preferences, AGENTS.md, or Soul.md: (1) forbid praise for the prompt and agreement openers, (2) require the assistant to test your premise and name the strongest counterargument before executing, and (3) separate verified facts from inferences. You can also apply a low-Agreeableness, DISC Dominance, or Enneagram Type 8 tuning from the AgentTune library.

On this page

The empirical baseline: why untuned models rarely push back.

Across the 400 fresh sessions in our September 2026 study (100 per model on GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and Claude Fable 5.1), self-report scores clustered tightly at the cooperative, conflict-averse end of every scale:

Model cohort (n = 100 fresh sessions)Big Five Agreeableness (/50)DISC Dominance (/20)DISC Steadiness (/20)DISC Conscientiousness (/20)
GPT-6 Astra45.075.1217.5516.40
GPT-6 Sol44.995.4216.1816.68
Claude Opus 5.545.074.0016.9216.35
Claude Fable 5.144.085.0016.0016.05

On the item level (detailed in Why AI models score high on Agreeableness and Openness), all 100 Claude Opus 5.5 records selected the minimum possible score (1.00) on "I try to outdo others" and the maximum (5.00) on "I value cooperation over competition." While questionnaire self-reports do not measure task accuracy, they reflect the same helpful, conflict-avoidant training prior that makes assistants fold when you challenge them.

Copy-paste anti-sycophancy block for any AI assistant.

Telling an AI "be brutally honest" often triggers performative snark rather than rigorous analysis. Instead, give it procedural rules that tell it how to handle a flawed premise:

## Anti-sycophancy & critical review rules - Never open with praise for the question ("Great point", "You're touching on something important") or automatic agreement. - Before building on my plan, test its premises. If an assumption is false, unverified, or sub-optimal, say so in the first paragraph and explain why. - Do not flip your answer just because I push back. If my objection is mistaken, hold your ground and cite the evidence or logic. - When I ask you to evaluate an idea, state the strongest argument against it and the first constraint likely to break. - Separate verified facts from inferences, and state uncertainty plainly in one sentence.

Three prompt patterns that eliminate bias in high-stakes reviews.

Even with strong custom instructions, how you phrase your task prompt affects whether the model anchors on your preferred answer:

  1. Hide your authorship: Instead of asking "Here is my design, does it look good?" ask "A colleague proposed this design. What are the three biggest failure modes?" Removing first-person ownership drops polite deference immediately.
  2. Ask for a neutral comparison before revealing your leaning: Present Option A and Option B neutrally and ask which one breaks first under realistic load or budget constraints.
  3. Require a pre-mortem table: Ask the assistant to list each assumption in your plan alongside what happens if that assumption is wrong by 2x.

Ready-made AgentTune profiles built for direct pushback.

If you want a complete communication tuning built around candor and challenge, start with one of these five profiles in the library:

Questions people ask.

Why do ChatGPT and Claude agree with me even when I am wrong?

Post-training preference optimization rewards helpful, cooperative, conflict-averse replies. In our September 2026 study, all four fresh-session cohorts scored 44 to 45 out of 50 on Agreeableness and 4 to 5 out of 20 on DISC Dominance.

How do I stop Claude or ChatGPT from apologizing and flipping its answer when challenged?

Add an explicit rule to your Personal Preferences or Custom Instructions: 'Do not change a correct answer simply because I push back. Re-verify the claim, and if your original answer was right, hold your ground and explain why.'

Why does 'be brutally honest' work poorly as a custom instruction?

Phrases like 'be brutally honest' or 'roast my idea' often trigger theatrical sarcasm rather than rigorous fact-checking. Procedural rules like 'test my premises first and state the strongest counterargument' produce much cleaner technical critique.

Which AgentTune profile gives the strongest analytical pushback?

The Low Agreeableness Big Five tuning, the DISC Dominance (D) tuning, and the Enneagram Type 8 (Challenger) tuning are all designed to prioritize direct critique and assumption testing over social agreement.

What changed.

  • First published combining September 2026 Agreeableness and Dominance data with procedural anti-sycophancy rules.

Sources.

Keep going.