Interesting breakdown. I’ve noticed the same pattern while experimenting with a few of these tools. They’re amazing for getting something working quickly, especially for prototypes or small utilities, but once you move toward production the real challenge becomes reliability and testing. The speed of generating code has definitely outpaced how we validate it. I’ve had cases where everything looked fine during development but weird edge cases started showing up once real users interacted with it. Feels like the workflow now is less about writing code and more about verifying what the AI produced. I’m curious too what tools or practices people are using to close that testing gap, because that seems like the next big bottleneck.

I’ve been experimenting with this a bit too. For small utilities and scripts it definitely speeds things up, especially when the requirements are clear. But once it touches an existing codebase or production systems, I still feel more comfortable writing most of it myself and just using AI for suggestions or quick scaffolding. The review part you mentioned is interesting though. When a big chunk of code is AI-generated, I think the focus shifts more toward understanding the logic and potential edge cases rather than just style or structure. Curious to see how teams eventually standardize this.