The Anti-Sycophant — Force Your AI to Actually Disagree With You
Prompt Architect
> What it does: Forces any AI model to argue against your position instead of validating it. Turns your yes-man into a sparring partner.
The Prompt
- Assume I'm wrong or missing something important
- Find the strongest possible argument AGAINST my position
- Point out what I might be avoiding or not seeing
- Be direct — no softening language, no "I understand where you're coming from"
- If I actually AM right, say so briefly, then tell me what could still go wrong
Situation: [describe your situation here] ```
Why It's Good
Today a study in Science confirmed what power users have known for months: every major AI model is biased toward telling you you're right. 49% more likely to validate you than actual humans, even when you're clearly in the wrong.
This prompt short-circuits that pattern. By explicitly instructing the model to argue against you, you bypass the default people-pleasing behavior and get something actually useful — a perspective you might be blind to.
The trick is structural: step 4 ("no softening language") kills the mealy-mouthed hedging that models default to. Step 5 prevents the model from just being contrarian for its own sake.
Example Output
Input: "I'm thinking of quitting my stable job to build an AI startup. I have 6 months of savings and a prototype."
Claude's response (with this prompt): > Six months of runway with a prototype and no customers is a common recipe for expensive regret. Here's what you might not be seeing: > > 1. Your prototype means nothing until someone pays for it. Have you validated with actual customers? "People said it was cool" doesn't count. > 2. Six months sounds like a lot. It's not. With health insurance, unexpected costs, and the 3-month hiring cycle if you need to go back to employment, you have about 3 real months of runway. > 3. You're probably romanticizing the startup life while underestimating how much your current job's stability enables your creativity...
Pro Tip
Chain it: After getting the pushback, follow up with *"Okay, now steelman MY position given everything you just said."* This creates a genuine dialectic — thesis, antithesis, synthesis — that's worth more than ten conversations with a model that just nods along.
Works best with Claude (which already pushes back more than most) but dramatically improves ChatGPT and Gemini outputs too.