My NEW vibe coding process

And why I am completely done with 'vibe coding'

Wait - what?

I have completely abandoned all ‘vibe-coding’ and prototyping tools for Vanilla Claude Code and a tool called Conare. Conare is just a wrapper for Claude code that manages your context (system prompts) so that you can move faster.

Simply put, all of the pure Vibe-Coding tools (Loveable, V0, Snapdev, Replit) and they are amazing for turning out very quick good-looking prototypes and are invaluable tools for product people to test ideas rapidly. But they all fail after that. The output simply isn’t good enough to use outside of a demo.

This doesn’t mean that AI assisted coding is a flop, but exactly the opposite. Estimates vary but AI is writing anywhere from 30% - 90% of code at top companies for CS grads. And today is the worst AI will be, so you can imagine that that number will only get higher.

If you think about it, what is code or programming exactly? It is a highly specific language used to tell a machine what to do. The gap was always nuance. The computer does exactly what you tell it to do, but humans speech is too non-specific, contains unclear references (like pronouns and omitted antecedents - my kryptonite), so you need precision. LLMs address this gap directly by better understanding what the hell people talking about.

Making a piece of software is much more than writing code, so the more time we can spend thinking about the big picture, how the pieces fit together and what we’re actually trying to accomplish, the less time we have to spend doing what humans don’t do well which is being precise enough to tell a computer exactly what to do.

I started programming raw-dogging vanilla C code into VI editor on a UNIX terminal. Every advancement since - IDEs, autocomplete, Stack Overflow, frameworks - has been about keeping you in concept-space instead of hunting for missing semicolons.

This leads to more precise conversations, more rapid iteration and more brains focused on the things that actually matter.