Abstract: Large language models (LLMs) now is challenged by cognitive overload in complex reasoning tasks due to its storage and processing of large amounts of information during reasoning. To address this issue, this paper introduces the self-guide method to guide LLMs to generate common sense knowledge and reasoning cues, allowing self-planning to improve performance. This method is integrated into the chain-of-thought (CoT) process, improving reasoning without requiring model fine-tuning or external tools. Experimental results demonstrate that the proposed method significantly outperforms baseline methods across four reasoning tasks and generalizes effectively in models with weaker reasoning capabilities.