Natural Language as the Bridge Between Pedagogy and AI: Paper 2 of 4

Paper 1 introduced a lightweight framework for encoding pedagogical expertise as discrete elements with natural language parameters, demonstrating how this structure enables constrained LLM generation of educational content at scale. This paper extends that architecture to audience tailoring, showing how the same pedagogy can produce fundamentally different learning experiences for different learners, without any change to the underlying pedagogical design.We introduce the concept of a ‘variant profile’, which is a structured set of learner-facing natural language rules that can be applied to a pedagogical content item to tailor its output for a specific individual or group. Related to this is our preferred methodology of creating a ‘primary variant’ and then using variant profiles to tailor this. Although not the main subject, this paper addresses the professional debate about how far tailoring should go, and argues that the framework makes that debate more tractable by making tailoring decisions explicit, testable, and reversible. Paper 3 extends the variant profile concept beyond text to visual and spatial design, covering deterministic and non-deterministic trade-offs in considerably more detail. Paper 4 extends the framework further to temporal sequencing and narration.Keywords: audience tailoring, variant profiles, differentiation, EAL learners, neurodiversity, constrained LLM generation, prompt engineering, educational technology