Improving Long-Horizon Imitation Through Instruction Prediction