I found I used fewer tokens by having many short prompts than long ones — because it spent many fewer tokens thinking and narrating.