Prompt-Guided Internal States for Hallucination Detection of Large Language Models