This is an application of the rule of three: you need at least three clients to write a good library, protocol, or other underlying abstraction. Except it proposes using a large language model to write the three clients and then throwing them away.