Can prompting reduce AI sycophancy or is it mostly model behavior?

I’ve noticed that Gemini often feels very agreeable in some conversations. Even when I ask for an objective opinion, it sometimes seems to validate my assumptions first instead of directly challenging them.

For example, when I ask whether my reasoning is flawed, it tends to respond with something like “That’s a valid concern” or “You’re making a good point” before giving criticism, which makes the criticism feel softened or less direct.

I’m curious whether this is something that can be meaningfully improved with prompts, such as asking the model to be more critical, or whether sycophancy is mostly a model/personality alignment issue. And I wonder if there are differences between Gemini, ChatGPT, Claude, etc. when it comes to disagreement or objective criticism.