Automatic code generation is crucial in modern software development, yet large language models struggle with real-world challenges like code versioning and multi-API invocation. Existing approaches, including direct generation and retrieval-augmented methods, often fail to ensure precise API usage. This paper introduces a simple yet effective two-step framework: rough code generation or retrieval followed by fine code editing. Experiments on VersiCode and BigCodeBench show significant performance gains in version-specific code completion and function-level programming. These results demonstrate the framework’s practicality in enhancing LLM-based code generation systems.