Large Language Models Can Self-Correct with Minimal Effort