I don’t know if anyone on the dev team realizes this, but the Gemini flash thinking—I don’t think it’s good. It feels like a to-do list rather than real reasoning.
It looks like you guys are trying to separate thinking and answering to achieve bigger context windows, because I feel the thinking and answer feel disconnected. You have to train the model to answer based on its thinking; you can’t let the model do it based only on your prompt. It won’t work because the model was never trained to use its thinking to answer.
Anyway, there are a lot of things to improve, but…