From Writing Traces to Personalised Support: Guiding LLMs with Stylometric Fingerprints

Abstract

Generative AI holds great potential for personalised learning, yet its outputs often remain generic. A key barrier is the difficulty Large Language Models (LLMs) face in interpreting quantitative learner data. This study introduces StyloPrompt, a three-phase method that converts quantitative writing analytics (stylometric fingerprints) into structured natural language descriptors to enable interpretable conditioning of LLMs. In Phase 1, variance analysis of over 1,000 student essays validated 19 robust features that stably distinguish authors. Phase 2 operationalised these fingerprints in a controlled style transfer task—serving as a proxy for adaptive educational communication—comparing raw..