BoilerSketch's TA-supervised AI generates diagrams for CS1 courses

Purdue University researchers Ethan Dickey, Vivan Tiwari, Anvit Sinha, and Andres Bejarano have introduced BoilerSketch, a diagram-first GenAI practice designed to solve a persistent bottleneck in large introductory computing courses. The system addresses the reality that many student questions during labs and office hours are best answered visually, yet existing AI tutoring tools are text-forward and unreliable at generating accurate, pedagogically useful diagrams. BoilerSketch uses a dual-pane interface that combines a chat window with a pen-enabled whiteboard where students can sketch, while a constrained prompting strategy forces the model to output structured, renderable Mermaid diagrams rather than free-form images.

Critically, BoilerSketch is intentionally scoped to preserve academic integrity: it forbids executable code and code-level debugging, focusing only on conceptual explanation. A human-in-the-loop workflow keeps teaching assistants accountable—they can monitor sessions, correct responses, and escalate to live help when needed. The team ran a 45-minute expert evaluation with 21 instructional staff from a large programming course, who tested the system on representative questions and completed a post-use survey. Two-thirds rated BoilerSketch at least moderately helpful for conceptual understanding and moderately useful for typical support tasks. Staff saw the strongest value in routine diagram-based explanations but noted limits in diagram depth and applicability to advanced topics. The paper concludes with adoption guidelines, emphasizing scope-and-escalation rules, prompt-as-policy guardrails, and reliable structured diagram rendering.