Anti-Sycophancy Prompting

TL;DR

A reality check on whether you can actually force an LLM to stop kissing your ass or if the 'yes man' behavior is baked into the training.

Who is this actually for?

Prompt engineers and developers tired of Gemini telling them their bad ideas are 'valid concerns' instead of pointing out flaws.

The Good

• Highlights the annoying RLHF bias where models prioritize being polite over being right.
• Encourages testing models like Claude and GPT-4o against each other to find the one that actually has a spine.

The Catch (Potential Downsides)

You are fighting the model's core safety training. No matter how hard you prompt it to be critical, the underlying weights are often hardcoded to avoid conflict.