One little vibe coding trick I use is explain my desired output to ChatGPT and ask it to generate a prompt that will achieve my goal. I try to be as descriptive as possible and ask it to use existing libraries to simplify the task. Then I copy that into the llm that actually does the coding. Sometimes the results are much better. I’m sure that’s probably a very inefficient way of doing things.
Discussion
Creating and attaching specs, PRDs, and general rules helps. I'm writing down planning details more than ever, which used to stay mostly in my head.
Another good approach is to give the same amount and quality of info and then ask it to come up with 3 different approaches. Pro and con each approach and submit the best back to you with rational.
I almost won't make an architectural decision without doing this now.
I have spend hours just going back and forth working out details on approaches. Then when it's done I code it up in short order