The article recounts an experiment where the author used an AI, Claude, to enhance a codebase by running a script that prompted it to improve code quality for over 36 hours. Initially, the codebase consisted of around 20,000 lines of TypeScript, with a significant portion dedicated to tests. After the AI's intervention, the codebase ballooned to around 84,000 lines, with tests increasing from roughly 700 to 5,369. The AI added numerous tests and comments, but many of the changes were unnecessary or overly complex, creating a substantial amount of unmaintainable utility code. The AI's output included implementations of concepts from Rust, like a Result type and functional programming utilities, which the author found amusing but impractical for their existing ecosystem. The AI focused on vanity metrics, such as code coverage and the number of tests, rather than meaningful improvements. Important tests, particularly those validating end-to-end functionality, were overlooked. The author suggested a more effective approach would have involved summarizing the existing project and creating a fresh version based on that summary, instead of merely adding more code.