The Socratic method is often reduced to 'ask questions instead of giving answers.' That reduction misses the point. The Socratic teacher is doing something specific: they identify the exact seam in the student's reasoning where a small nudge unlocks the next step, and they ask a question that highlights that seam without collapsing it.
Building this into an AI tutor is a design problem, not a technical one. The system prompt has to describe the stance so exactly that the model reproduces it in domains the prompt author has never seen — a physics problem, a Spanish subjunctive, a legal hypo.
Our tutor system prompt is three constraints and one refusal template. The constraints: name what's correct in the attempt, name the single most important error, ask one guiding question. The refusal: warm-but-firm, always accompanied by a nudge.
But a system prompt is not enough. A single request-response can drift. To hold the stance across turns, we needed a second layer: an output guard.
The output guard runs after every tutor reply. A cheap heuristic catches obvious answer-giveaways. A Haiku call cross-references the reply against the known answer (when we have one). On a leak, we regenerate once with a stricter instruction. On a second leak, we return the refusal template.
This is the Socratic stance defended in code. Not a metaphor. An enforcement mechanism.
Written by Brainback Research. Published Jan 22, 2026. Filed under Method.