Siddharth Saoji 
When you sit down to write an article, you don’t start with a spec. You write a first draft rough, wrong in places and then you refine it, and refine it again and again till you are happy with it. The final piece is usually unrecognisable from where you started, and that’s the point. The first draft isn’t the destination. It’s the thing that breaks the inertia. I think vibe coding with AI coding agents should work the same way. Instead of investing heavily in a spec upfront, stumble toward a first working MVP. However rough, however far from ideal. Then refactor. And refactor again. Yes, this means more tokens, more time to production, more context management. But it also means deeper product understanding, and more genuine engagement in the process of building.
“But what about planning, then iterating?”
The obvious pushback: write a solid spec, then refine recursively within it. Best of both worlds. But there’s a meaningful difference between planning then building, and building then discovering. When you plan first, you commit to a vision before you’ve felt the product. A spec becomes a contract with your past self, the person who imagined the thing before seeing it move. But products reveal information that plans can’t. A clunky flow exposes a deeper UX problem. An awkward constraint opens a better solution. These signals only reach you while building, not planning. Then there’s the inertia problem. Planning is heavy cognitive load without seeing something tangible for the initial effort. The draft, by contrast lowers the barrier, creates momentum, and makes the next step obvious.
“Even the feedback loop is iterative”
There’s one more parallel worth noting. When you revise an article, you reread until something feels off or right, no checklist, just a signal. With agentic coding, you run outputs, watch what breaks, adjust. Neither process has a clean objective metric upfront. Both rely on iterative assessment.
“Why this convergence is happening now”
This shift is possible because of how generative AI works. To an LLM, there’s no fundamental difference between generating an article and generating code, both are next-token prediction. The same dynamic that’s always defined good writing is now applicable to software development, the differentiator isn’t whether you can build, it’s how tasteful and elegant an experience you can provide.
The question worth sitting with isn’t whether this approach is better. It’s whether you’re already doing it instinctively and what might happen if you did it more deliberately.