Stepping into SDD (Spec-Driven Development)

While this may not be an early move, I’ve finally stepped into SDD (Spec-Driven Development).

I’ve been hearing for quite some time about the idea of aggressively delegating coding to AI. I even ran a few experiments myself. Still, I believed that fully entrusting a large project to AI wasn’t quite realistic yet. I knew there were people attempting it—and even some showing solid results—but, to be honest, I didn’t have the mental space to explore it at the depth it deserved. I tried various forms of so-called “vibe coding” and confirmed that they could work to some extent. However, there was one critical piece missing if this approach were to be applied to real services: consistency of output. With carefully reused prompts, you could achieve a certain level of consistency. That might be acceptable for a personal project, but shaping a team-wide way of working is a very different challenge—and one that didn’t seem easy to solve.

For those reasons, I more or less set aside the idea of AI-driven coding strategies—until a few weeks ago, when I came across SDD. The entry point was encountering the keywords BMAD and Speckit. After spending some time learning about them, it became clear that these were simply practical methodologies for applying SDD in real-world scenarios. The core value was SDD itself. The more I explored it, the more it felt genuinely feasible. To validate that intuition, I tested it on a personal project for about one to two weeks, and that initial intuition gradually turned into confidence. I also gained a reasonable understanding of how SDD could be applied not only to projects built from scratch but also to existing, already-set-up codebases. Of course, once applied in the real world, countless edge cases will surface—ones I haven’t anticipated yet—and those gaps will need to be filled. Still, that’s a solvable problem. With approval secured for the necessary resources, it felt like the right time to move forward.

After formally declaring this direction within the teams, I started by applying SDD to a small repository as a personal exercise. That’s when the first real issue surfaced: I hit the Claude Pro Plan token limit. Claude resets usage every five hours, and fortunately, I happened to hit the limit right near the end of a usage window. I waited, worked on something else, and resumed as soon as the tokens refreshed—thankfully, with the context still intact. That relief didn’t last long. About an hour later, I hit the limit again. And that’s how—partly out of frustration—I ended up subscribing to the Claude Max Plan.

For now, I’m running tests under my personal account, but for team use, I’ve already instructed our purchasing owner to proceed with a Team plan subscription. Even so, I can see token limits becoming an issue again before long. But that’s a problem for another day. From here on out, I suspect I’ll have more than enough challenges to keep me busy.