The biggest mistake people make when vibecoding is assuming that AI knows the proper way to build things without validation and that AI thinks of all of the edge cases, so that you do not have to. The unfortunate truth is that AI only thinks about the context that you give it. So, if you don’t bring up things like security, the fact that you’re building something multi-user, the fact that you’ll eventually have X number of users using the platform, you fall victim to AI making errors in implementation because it’s misinformed.
The ultimate fix for this from my perspective would be that we identify a system prompt that can make AI think about all of these things and ask you a bunch of questions to get to the answers like security etc that you may not have thought of, then we pair that with a set of “vibe-coded approved components”, built by engineers to help vibe coders follow best practices (i.e. the right performant way to paginate a list).
In the short term, while that silver bullet doesn’t really exist, or is at least sort of the way there in a bunch of different places, use plan mode often, and in the early days of building your systems, consult engineers you know to make sure AI is going down the right path and then build upon them.
