If you use open models you can also read the thinking tokens, I think this makes a huge difference because one tends to read them leading to better understanding of the end results and also giving opportunities to correct course if its hung up about ambiguity and stuff like that. It also helps a lot with not being distracted while the clanker is working, instead of going to youtube or reading up on whatever, you spend more time with the code and the thoughts behind it.
It stops the constant context switching.