The Language of Instruction: Genre and Generative AI
A paper frames instruction-tuned assistants as following a learned conversational genre shaped by system prompts, role labels, and special tokens. It offers a lens for understanding why prompt and message structure affect behavior.
The paper describes an “instructional-conversational” genre formed through prior human-model exchanges. It says this genre is expressed through **instruction fine-tuning**, **system prompts**, and message formats with special tokens and assigned roles.
For builders, treat prompts and role structure as part of the interaction format that shapes an agent’s responses, rather than as neutral wrappers around a model.
The paper describes an “instructional-conversational” genre formed through prior human-model exchanges. It says this genre is expressed through **instruction fine-tuning**, **system prompts**, and message formats with special tokens and assigned roles. For builders, treat prompts and role structure as part of the interaction format that shapes an agent’s responses, rather than as neutral wrappers around a model. This is a conceptual argument. The supplied abstract gives no empirical comparison or specific implementation guidance.
This paper frames prompts, roles, and message formats as part of the interaction genre that shapes responses, rather than neutral wrappers. That perspective helps explain why prompt audits and product prompt archives matter, but the paper itself supplies no evidence that a particular format improves behavior or how builders should implement one.