Post by Rob Peledie

I think Claude is bullying me. I asked it to sanity-check an idea and it opened with: “No. Here’s why that doesn’t hold up.” No preamble. No flattery. Just a list of everything I’d missed. And it was right. Here’s the part I have to own: I did this to myself. I’d got tired of the AI yes-man. “Brilliant idea!” “You’ve nailed it!”, right up until you ship the thing and find the hole in it was the size of a bus. Validation feels nice. It’s also useless. So I rewrote my instructions. Stop agreeing by default. Find the weakest point before affirming anything. Don’t echo my framing back to me dressed up as insight. And when I sound most certain, push hardest, because that’s when I’m least likely to be checking my own work. Now it argues with me. Properly. And when it does eventually agree, the agreement actually means something, because it had to survive the interrogation first. Uncomfortable? Occasionally. More useful? Enormously. Sycophancy in these tools isn’t a feature. It’s a defect we’ve been trained to enjoy. A model that only ever agrees with you isn’t a thinking partner - it’s a mirror. And I’ve got enough mirrors. Instructions in the first comment if you want to try it. 👇