I'm a senior engineer doing React development since 2015, so I'm capable of writing the code myself. I've found, that I'm more productive if I guide this process using English from a middle to high level of guidance. I frequently mention specific names/properties/commands in my prompts. I mostly do bite-size queries. However, I almost don't touch the source code with bare hands anymore.
As I discussed this topic with friends, I've met people with polar experiences. Some of my friends who need to do web apps using React-TS are very happy with vibe coding. They report productivity boost. Some people even report ability to do things they couldn't do otherwise: desingers, product managers, backend engineers.
However, I've also heard of many negative experiences. People who write for Android, iOS, Telegram bots, Rust, even some Python scripts, report small to zero productivity boost and report uselessness of current Cursor-like experience.
I suspect that's because even Sonnet-3.7 is weak for languages and frameworks which aren't top-1 (top-3?).
If you've tried vibe coding for the last 6 months, can you mention your tech stack and whether your experience was positive or negative?
Another piece of magic sauce: waste the context. Spend the tokens as if you’re partying last time in your life. Throw anything remotely relevant at it.
I haven’t yet found piece of context that made it perform worse. Most of the time the opposite happens - it finds stuff that I haven’t noticed in the big chunks of data I threw at it (error logs, documentation, stack traces)
I remember when I was just learning programming I had a mental block around nested for-loops. I couldn’t trace more than 1000 operations in my head, so I thought that computer wouldn’t be able to as well. To my surprise a computer can easily handle a million operations in a split second.
With AI is a similar feeling, I feel constrained by the amount of my own “RAM”, but once I let go, it almost always surprises me.
The corresponding commit message was 46 neatly wrapped lines, not counting blanks or the commit header, including a bunch of notes that will probably eventually make their way into documentation.
I didn't try to involve AI in this, and I strongly suspect it wouldn't help. Coming up with the ideas and putting them in precise language isn't really any easier for me than actually writing the code. While writing it, lots of things didn't work the first time, but I never felt "stuck".