Persona Framing in Human–LLM Interaction

Large Language Models (LLMs) exhibit a high degree of stylistic adaptability, yet they frequently produce unstable responses in terms of discursive register, enunciative position, and pragmatic coherence across extended interactions. A common practice among users is persona prompting, in which the model is instructed to assume a specific discursive role. Although widely used in practice, this technique has been treated primarily as a prompt engineering strategy, with limited analysis from the perspective of linguistics and discourse pragmatics. This paper proposes interpreting persona prompting as a mechanism of discursive framing (persona framing) in human–LLM interaction. It argues that persona instructions function as pragmatic devices that establish discursive registers, delimit enunciative positions, and constrain the stylistic space of generated responses. At the same time, the concept of persona framing is proposed not only as a linguistic interpretation of an existing technique, but also as a useful conceptual lens for the design of structured interactions with language models. As an exploratory study, we analyze structured interactions with a system of discursive personas called Nemosine, in which different linguistic roles have defined discursive vocations and are maintained throughout dialogue. Based on this analysis, we propose a preliminary typology of framing levels in human–LLM interaction: (1) ephemeral persona framing, in which the persona appears only in a single isolated turn; (2) persistent persona framing, in which the persona remains active across multiple turns; (3) environment framing, in which interaction takes place within a symbolic space of interpretation that allows the invocation of different personas; and (4) system-level persona framing, in which the architecture of the interaction system is organized around multiple personas with specific discursive functions. The qualitative analysis suggests that persona-based framing contributes to greater consistency of register, greater stylistic stability, and greater pragmatic predictability in the model’s responses. The results indicate that persona framing may be understood as an emerging linguistic interface for interacting with LLMs, connecting prompt engineering practices with classical theories of discursive framing, polyphony, and pragmatic organization of discourse.