When GPT-6 Astra was released, I panicked. I’d been thinking about AI’s impact on software development for a long time. GPT-6 Astra was the first model that made those questions stop feeling theoretical and start feeling real. The programming skills I’d spent years developing suddenly felt obsolete. Astra was so fast, so capable, and so productive that writing code by hand suddenly felt pointless.
After spending a few days with the new model, my perspective started to change. Instead of asking whether AI was replacing me, I began asking how to adapt to this new world. The challenge wasn’t learning another programming language or framework. It was learning an entirely new way of working.
Two days ago, I spent most of my development session trying to figure out that workflow. I experimented with prompts, learned the ChatGPT desktop app, and spent a lot of time thinking about how to break work into more explicit tasks. Over the years, I’d developed a workflow where most of the implementation details stayed in my head. A short to-do item was enough to remind me of an entire design. AI can’t read my mind, so I have to externalize that thinking. Planning suddenly matters much more than it used to. During that session, I asked Astra to implement a cutscene for my game. It completed the task remarkably quickly.
Yesterday, I didn’t ask it to generate any new code. I spent my entire session reviewing what it had produced the day before. I spent well over an hour reading through the code, checking the architecture, looking for mistakes, and verifying that everything made sense. Surprisingly, I liked what I saw. Previous generations of AI models often made architectural mistakes. Fixing them usually meant throwing away entire sections and rewriting them from scratch. Astra feels different. The overall design was solid. Most of the issues were relatively minor. Looking back, a lot of them were my fault. I’d forgotten to include my project’s coding standards in the prompt, so the generated code wasn’t always consistent with the rest of the codebase.
What surprised me most wasn’t the quality of the code, it was how boring the process had become. Instead of spending the afternoon designing systems and writing code, I spent it reviewing someone else’s work. Maybe it was more productive, but it was also incredibly boring. I’m still early in this transition, and I don’t have answers yet. I suspect software development is becoming less about typing code and more about decomposition, planning, communication, and critical review. That’s a very different skill set.
The question I’m trying to answer now isn’t whether AI will make me more productive. It’s how to make this new way of building software feel as rewarding as the old one.