In Gemini, it is really good to have large window with 1M tokens. However, around 100,000 it starts to make mistakes and refactor its own code.
Sometimes it is good to start new chat or switch to Claude.
And it really helps to be very precise with wording of specification what you want to achieve. Or repeat it sometimes with some added request lines.
GIGO in reality :)
Oh my, I hate it when it rewrites >1k LOC. I have to instruct it to "modify only ..., do not touch the rest" and so forth, but GPT does not listen to this often, Claude does. I dunno about Gemini.