Effective Prompt Extraction from Language Models