We've had success in various applications using a semi-structured natural language for LLMs. Structured inputs, such as a JSON document, are clear and precise and give the model an indication of the type of response being sought. Constraining the response in this way helps narrow the problem space and can produce more accurate completions, particularly when the structure conforms to a domain-specific language (DSL) whose syntax or schema is provided to the model. We've also found that augmenting the structured input with natural language comments or notations produces a better response than either natural language or structured input alone. Typically, natural language is simply interspersed with structured content when constructing the prompt. As with many LLM behaviors, we don't know exactly why this works, but our experience shows that putting natural language comments in human-written code also improves the quality of output for LLM-based coding assistants.
