Where decades of learning-science research meets AI tools — intelligent tutoring systems, cognitive tutors, worked-example sequences, productive failure, AI-generated feedback, and the formative-assessment loops that hold it together. We curate hard here: the classical learning-science principles (Sweller, VanLehn, Kapur, Renkl, Black & Wiliam) have strong meta-analytic evidence; the AI-specific re-implementations often inherit that evidence, but not always — and where the AI angle outruns the research, we say so. TPACK is included as the framework teachers need to put any of this in motion.
Design cognitive-tutor architecture — ACT-R cognitive model, knowledge-tracing of student state, mastery thresholds, step-level hints and error messages — so the tutor actually knows what the student knows and adapts, instead of just delivering branching worksheets.
Design tutorial dialogue — step-based interaction, ICAP-aligned prompts (Interactive > Constructive > Active > Passive), mixed-initiative turns, and AutoTutor-style follow-up moves — so the tutor pulls reasoning out of the student instead of dumping explanations on them.
Design hint hierarchies that fade from vague (point me at the right area) to specific (here's the next step), released only when the student is genuinely stuck, so the system supports productive struggle instead of bailing them out at the first wrong answer.
Design AI-generated feedback so it actually moves learning — specific, focused on the task and process (not the person), tied to clear criteria, and actionable. Built on Shute/Narciss/Hattie classical feedback principles, applied to LLM and automated-feedback systems.
Design problem-solving experiences that come before instruction — letting students grapple with a hard problem, generate partially-correct approaches, then receive instruction that consolidates what they have just struggled with. Kapur's PS-I (problem-solving followed by instruction), built on Bjork's desirable difficulties tradition.
Design when and how to move students from studying worked examples to solving problems on their own — anchored in the expertise-reversal effect. Keep novices in high-assistance mode (full worked examples → completion problems → faded steps), move them to independent problem-solving as prior knowledge grows, and use diagnostic info to detect the crossover instead of assuming everyone is at the same place.
Design AI-supported group tasks that structure interaction, balance participation, and surface the regulation moves that make CSCL actually work — building on the Dillenbourg / Järvelä CSCL tradition and Slavin's cooperative-learning principles, with AI used to prompt, mirror, and scaffold the group rather than replace it.
Design the inner and outer assessment loops that make AI-mediated instruction actually responsive — each student response feeds into the next instructional move within a task (inner loop: VanLehn) and into the next task in the sequence (outer loop), built on the Black & Wiliam formative-assessment tradition so the AI is doing assessment-for-learning, not just delivering more content.
Work through the TPACK lens (Mishra & Koehler) for a specific technology or AI tool — what pedagogy the tool affords, what content knowledge it carries assumptions about, and where teacher knowledge has to bridge the three. Use when adopting an AI tool, reviewing ed-tech, or planning a tech-integration cycle so the integration is pedagogically aligned, not just procurement-driven.
Read LMS and quiz-platform learning-analytics dashboards and translate the patterns into teaching decisions — which students need re-teaching, which concepts the class collectively missed, which engagement signals are real versus noise. Anchored on the most recent meta-analysis: dashboards produce an around-average effect on academic performance (g=0.36), with the effect concentrated in higher education (K-12 effects are weaker) and in knowledge-acquisition outcomes (engagement and SEL outcomes show much smaller effects). Most useful when the dashboard data feeds directly into the next lesson, not just into report-card endpoints.
Heard a claim somewhere else? Type it here — we'll build a prompt you can paste into ChatGPT, Claude, or any chatbot. The prompt asks the model to cite real meta-analyses, separate strong from weak evidence, name boundary conditions, flag exaggerations, and discuss cost-effectiveness.