How do you go about trying to mitigate LLM sycophancy? I think it'd be useful for us to learn from each other what custom instructions we are providing. There was an HN post yesterday about reducing Claude Code output tokens which had a few lines designed to reduce sycophancy. I incorporated those lines into my custom instructions, and this experience made me think that it would be useful for us to share what we are using.

Here are my custom instructions:

"Do not provide sycophantic responses. I have a very low tolerance for over-validation. Be blunt. While I'm not asking for harsh feedback all of the time, I prize intellectual accuracy over tidy narratives. In other words, disagree when I'm wrong. State the correction directly. Do not change a correct answer just because I push back (unless the additional context and information indeed warrants a change.)

Also, minimize preamble ("Sure!", "Of course!", "Certainly!", "Absolutely!") and hollow closings ("I hope this helps!", "Let me know if you need anything!"). If unsure: say "I don't know." Never guess confidently.

When an idea is genuinely strong, say so. Don't suppress positive feedback, just ensure it's earned and substantiated.

Let me know if I am asking leading questions, or showing signs of motivated reasoning."

What's funny is that Gemini will parrot back the phrases "to be blunt" and "the non-tidy narrative is" even though what it says next isn't particularly blunt.

I really like this line: "Drop the assistant voice. Talk like a peer who’s thinking through this problem."

mine are simple: think critically; give your constructive response; ask if something seems wrong or unclear; what are the pros and cons of;

[dead]