Why Are Experiences of Vibe Coding So Polarised? (www.stephen-cresswell.com)

A recent discussion on Jason Gorman's blog highlights the polarized experiences developers have with Generative AI-assisted software development, particularly with Claude Code. While some engineers report productivity boosts of up to 20 times, others face issues like excessive code generation, architectural drift, and operational risks. The key to these divergent outcomes isn't necessarily the tools themselves but rather how developers apply them, framed by their individual goals and constraints. Gorman emphasizes that using Generative AI effectively requires explicit optimization targets, such as minimizing operational debt and achieving clean, malleable code. A controlled experiment testing three different development styles on a URL shortener service demonstrates the nuances of "vibe coding." The study found that a structured approach incorporating explicit prompts and manual reviews produced cleaner code with lower operational debt, while an unattended method, though faster, resulted in overwhelming complexity and unreliable outcomes. Gorman concludes that successful deployment of Generative AI hinges less on the decision to adopt the technology and more on how organizations implement it to align with their long-term engineering objectives. This underscores the need for more controlled methods to better harness AI's capabilities in software development.