It's not perfect, but on average, it should improve results significantly compared to using models individually. If anyone wants to improve it, there a lot of gains to be made by adding context about the strengths/weaknesses of each model in the final prompt.

→ View original post on X — @mattshumer_