Maybe have the prompt tell the model emit the model "level" needed between 1 and 10 for each subtask it generates