I've been experimenting with "vibe coding" tools like Cursor and Claude Code (3.5)—the kind of AI pair programmers that promise to understand your code, suggest fixes, and help you fly through bugs without touching the keyboard too much. In theory, it's magical.
In practice? Well... let me tell you a story.
Embarking on the journey of translating C++ code into TypeScript, I enlisted the assistance of Claude 3.5 Sonnet, Anthropic's AI language model. My objective was to achieve a faithful, line-by-line translation of Config.hpp and Config.cpp, ensuring type safety by utilizing TypeScript types that closely mirror the original C++ structs and classes. The aim was to preserve the original file and class structure.
Ever been deep into a vibe coding session, grooving along with your AI assistant, only to have your flow completely disrupted because your AI buddy stubbornly refuses to admit it's made a mistake? Yep, same. Today, let's dive into the quirky world of vibe coding and unravel one of its biggest headaches—LLMs' pesky self-reflection limitations.