Filippo's article recounts his experience using Claude Code to debug a Go implementation of the ML-DSA, a post-quantum signature algorithm. After spending time on the implementation, he encountered persistent issues with signature verification, leading to failed tests. Despite initial frustration, he decided to leverage Claude Code, which quickly identified a complex bug. The AI discovered that he had improperly merged functions that handle high bits, causing the verification process to fail. This intervention saved him considerable debugging time, allowing him to refactor the code more effectively. In a subsequent testing phase, Filippo faced two additional bugs in the signing implementation. The first was a miscalculation of constants in the Montgomery domain, which took him hours to uncover. Claude Code, however, was able to identify this error much faster. The second bug involved an incorrectly sized encoded value in the signature. Filippo found it interesting that the AI struggled more with the second, seemingly simpler issue, possibly due to the nature of the output it was processing. Nevertheless, Claude's ability to pinpoint these bugs without needing to trust its suggestions for fixes demonstrated its utility as a debugging tool. Filippo reflects on the potential for AI tools in programming, particularly in debugging low-level cryptography code. He envisions a future where AI could automatically diagnose issues as they arise from failed tests, streamlining the debugging process even further. This case study not only illustrates the effectiveness of Claude Code in his work but also highlights the ongoing evolution of AI as a valuable tool for developers.