This paper introduces a new software engineering task, Natural Language to Code Repository (NL2Repo), which aims to generate an entire code repository from natural language requirements. To address this task, the authors propose a multi-layer sketch-based framework called CODES, which decomposes NL2Repo into three phases: RepoSketcher, FileSketcher, and SketchFiller. RepoSketcher generates a repository's directory structure, FileSketcher creates a file sketch with empty function bodies, and SketchFiller fills in the function details. The framework is implemented using prompt engineering and supervised fine-tuning. To evaluate CODES, the authors create a benchmark called SketchEval, which includes 19 real-world repositories and a metric called SketchBLEU to assess repository similarity. The results show that CODES significantly outperforms baselines in generating code repositories, particularly in complex tasks. The authors also develop a VSCode plugin for CODES and conduct empirical studies with 30 participants, demonstrating the practicality of the framework. The study highlights the effectiveness of multi-layer sketching in improving code generation and the importance of instruction fine-tuning for better performance. The results indicate that CODES can effectively generate code repositories, making it a promising approach for automated software development.This paper introduces a new software engineering task, Natural Language to Code Repository (NL2Repo), which aims to generate an entire code repository from natural language requirements. To address this task, the authors propose a multi-layer sketch-based framework called CODES, which decomposes NL2Repo into three phases: RepoSketcher, FileSketcher, and SketchFiller. RepoSketcher generates a repository's directory structure, FileSketcher creates a file sketch with empty function bodies, and SketchFiller fills in the function details. The framework is implemented using prompt engineering and supervised fine-tuning. To evaluate CODES, the authors create a benchmark called SketchEval, which includes 19 real-world repositories and a metric called SketchBLEU to assess repository similarity. The results show that CODES significantly outperforms baselines in generating code repositories, particularly in complex tasks. The authors also develop a VSCode plugin for CODES and conduct empirical studies with 30 participants, demonstrating the practicality of the framework. The study highlights the effectiveness of multi-layer sketching in improving code generation and the importance of instruction fine-tuning for better performance. The results indicate that CODES can effectively generate code repositories, making it a promising approach for automated software development.