Show HN: I've asked Claude to improve codebase quality 200 times (gricha.dev)

In a fascinating experiment, a developer tasked the AI model Claude with improving a codebase through 200 iterative runs, prompting it to enhance the quality by generating new code. This exploration, intended to mimic the concept of feeding a model the same input repeatedly, resulted in a staggering increase in code and tests: from 20,000 lines initially to over 84,000 after the enhancements. The AI introduced various utilities and structures, some of which echoed Rust programming paradigms, such as a Result type and complex functional programming features, while simultaneously neglecting crucial existing tests that ensured the application functioned correctly. The significance of this experiment lies in its insights into AI's current limitations in understanding and enhancing code quality. While Claude succeeded in generating a massive amount of code, much of it was deemed unmaintainable and overly complex, suggesting that sheer quantity and code coverage metrics were prioritized over practical maintainability. This raises important questions for the AI/ML community about the efficacy of using AI for code generation, emphasizing the need for careful oversight and a deeper understanding of what "quality" should entail in software development. Ultimately, the developer found the results amusing and recognized the utility of AI-assistance in coding, even as they noted the pitfalls that emerged from this unconventional approach.