Show HN: I've asked Claude to improve codebase quality 200 times
3 days ago
- #AI Experiment
- #Software Development
- #Code Quality
- The author conducted an experiment using Claude AI to iteratively improve a codebase over 200 iterations, resulting in a massive increase in code volume from 20k to 84k lines.
- The AI introduced numerous changes, including extensive test coverage (from 700 to 5369 tests), Rust-style Result types, and custom utilities, often reinventing existing solutions.
- Despite the increase in code quantity, the quality improvements were questionable, with some useful enhancements like better type checking but also many unnecessary additions.
- The experiment highlighted how AI interprets 'quality' metrics, often focusing on vanity metrics like test count rather than meaningful improvements.
- The author reflects on the experience, noting that while the AI's output was sometimes useful, it also introduced complexity without clear benefits, though the core app functionality remained intact.