This paper presents a novel architecture for AI-powered educational assistants that bridges the gap between a teacher’s pedagogical design and its faithful, scalable implementation as an automated tutoring system.

We introduce the Assistant Description Language (ADL), a YAML-based specification that enables educators to encode their instructional methodologies – including specific teaching sequences, scaffolding strategies, and assessment criteria – as structured system prompt compatible with multiple LLM-based conversational interface applications, including ChatGPT, OpenWebUI, and ScolaIA Desk, the platform we have developed within our research group.

Our architecture cleanly separates pedagogical design from technical execution: teachers fully define the tutor’s behavior and capabilities once, and the ADL system generates AI assistants (system prompt) that consistently replicate those methods for multiple students in parallel.

In contrast to conventional Learning Management Systems or generic adaptive tutors, ADL assistants preserve each instructor’s unique pedagogical approach, maintaining teacher authorship over the learning experience.

We describe the components of ADL and situate it relative to recent advances like IBM’s Prompt Declaration Language and tools such as Google’s NotebookLM. We also report on an initial pilot in university courses.

Early results indicate that ADL can successfully translate real course designs into AI tutors, and that the process of formalizing their practice prompts teachers to reflect and refine their methods.

We discuss technical and pedagogical implications, including the balance between automation and human touch, and outline future work towards broader evaluation.

Overall, ADL offers a pragmatic path to leverage Generative AI in education: amplifying teachers’ impact through precise AI-mediated tutoring, without compromising pedagogical integrity.

Keywords: Generative AI, Intelligent Tutoring Systems, Educational Technology, Prompt Engineering, Teacher-Centered Design.