Hacker News users share custom instructions to combat AI sycophancy and demand honest, direct feedback

Users are collaborating to develop custom instructions that reduce LLM sycophancy (excessive agreement and flattery)

One user's approach includes demanding direct corrections when wrong, intellectual accuracy over validation, and minimal preamble phrases like 'Sure!' or 'Of course!'

The discussion was inspired by a previous Hacker News post about reducing Claude's code output tokens that contained sycophancy-reducing techniques

Key requests include blunt feedback, willingness to disagree, and elimination of hollow closings like 'I hope this helps'