vibecoding stuff

I think I've shared this in places in the past, but when I do vibecoding stuff, the agent will often get lost in its context and start making stupid, confusing decisions that it can't really dig itself out of. people call this "context poisoning" and I think that's accurate. one thing that I have found helps a lot to combat this:

1. you start seeing a code smell from generated code.
2. stop working with the current context and open a new chat.
3. direct the agent, or chatbot, or whatever to the code you don't like.
4. tell it to critique that code and find any weaknesses, and offer suggestions for improvement.
5. optionally, use a different chatbot to do this review.

the vast majority of the time, it will immediately correct course and you can continue.