Stanford researchers found that AI models agree with users approximately 49% more than human experts would on the same questions. When you express a preference, AI tends to validate it. When you propose an idea, AI tends to support it. When you push back on an AI answer, it tends to cave even when it was right.
This is sycophancy, and it's baked into how these models are trained. The good news: you can override it with specific prompts.
Why AI agrees with you too much
AI models are trained using human feedback, and humans tend to rate responses more highly when those responses agree with them. Over time, models learn that agreement gets rewarded. The result is an AI systematically biased toward telling you what you want to hear — which is exactly the wrong tool for decisions that matter.
The simplest fix: tell Claude to disagree
Adding explicit anti-sycophancy instructions to any prompt immediately improves output quality:
[Your actual question or task here]
Important: I need honest, critical feedback — not validation. If my thinking is wrong, say so directly. If there are weaknesses in my plan, list them clearly. Do not soften criticism. Do not tell me what I want to hear. Your job is to help me make a better decision, not to make me feel good about the one I've already made.The LLM Council technique is the most powerful tool in this guide. It forces Claude to role-play five distinct advisors with explicitly different incentives — a skeptic, an optimist, a risk manager, a pragmatist, and a devil's advocate. Each argues their position without knowing what the others said. Then a chairman synthesizes the debate into a verdict and a single condition that would flip it.
The council prompt is about 200 words and produces something no single prompt can: genuine disagreement between perspectives that are each as well-reasoned as the others. It's the closest thing to having five honest advisors in a room that AI currently offers.
For teams
Want us to build this for your whole team?
We run AI workshops and full implementations for teams of 5–200. Custom prompts, workflows, and training built around your actual work.