System prompt changes help, but ultimately they don't do a great job of overcoming the model's learned tendencies in my experience.