We propose an information-theoretical metric that helps determine the optimal order of demonstrations for in-context learning in large language models.