Thank you for your input.
I actually started with a very similar approach using ChatGPT for a python study plan. I think it is a great learning tool and helped me a lot, especially because of exercises and the 1:1 teaching where you can deep dive what ever interests you the most right now.
I am sure I would have learned much more with the current version of ChatGPT (not even special learning AI) and free time instead of school and even university... But that's another topic.
I had a project (finance tracker and planer) I started coding after some lessons.
I later shifted to let ChatGPT write most of the code and make sure I understand all of it and look at the used Syntax etc. There are a lot of logical problems it can't solve in its own, especially when the project gets more complex, so you have to solve these on your own, but for the coding it worked surprisingly well, is fun and keeps me more motivated because I see more rapid results. I used python and streamlit for a web based program I can access through a VPN and also made a small flutter app to save cash transactions.
I agree with your two vibecoding improvement tips and think those are best learned by doing projects and experimenting.
But a third point would be tools. With the rapid evolution of AI I guess there are better AIs for vibecodeing or certain prompts and tricks that streamline the experience. I simply have no overview and am searching for a better entrance in that rabbit hole.