What works for me is:

- Use a fast model like DeepSeek Flash V4 on high (it's Sonnet level, but fast and cheap).

- While the LLM is working, start writting your next prompt. A good prompt usually takes between 1 and 10 minutes to write anyway.

Doing this should keep you busy enough to never leave flow.

But it is intense and demanding when the LLM is fast, I'll tell you that.