Chekos' team
I expected to build a lesson planner.
Core trick
Stores learner profiles and constraints as primitives, then composes stage-directed lesson plans calibrated to each group and setting.
Tags
Description
I expected to build a lesson planner. The first version was done in thirty minutes. So I kept going.
Maestro is a pedagogical reasoning engine built on five primitives: skills, dependencies, learner profiles, groups, and constraints. It interviews the educator, then composes stage-directed plans calibrated to the students, setting, and teaching style. A trail walk on ethnobotany for third graders with no Wi-Fi. An evening data viz workshop for tired professionals. Same engine. Two educators teaching the same topic get different plans — a lesson that doesn't fit the person delivering it won't work.
It assesses through conversation using SOLO Taxonomy. It learns every session. Students share learner profiles with parents in Spanish or employers in English.
Asked to create slides for AP Spanish Lit, the agent installed LibreOffice on its own, QA'd its output, and themed the design to Magical Realism — no instruction existed for any of it. That's Opus 4.6 with the right primitives.
chekos/pedagogical-engine★ 00 forksTypeScriptNOASSERTIONCreated Feb 12, 2026 · last push Feb 17, 2026
README excerpt
# Pedagogical Engine **An AI teaching partner that reasons about pedagogy — then argues with you about it.** Built with Claude Opus 4.6 and the Claude Agent SDK for the Cerebral Valley x Anthropic Hackathon (Feb 2026). --- ## The story If a student can run a pandas groupby, they probably already know how to filter data, declare variables, and open a terminal. Test the high-level skill, infer the rest. That's dependency inference, and it makes assessment 3-5x faster. That was the starting point. Then things got weird. The engine learned to simulate lessons before they happen — running friction analysis against actual student profiles. It'll tell you that Sofia will get stuck at minute 22 because her data cleaning skills are at Remember level, not Apply. Then it learned to argue. Ask for a 30-minute lecture for beginners and it'll tell you why that's a bad idea, citing Marcus's skill gaps and Alex's prerequisite holes. Not generic warnings. Your students, your data. Then it learned to explain itself. "Why did you pair Marcus with Sofia?" returns the actual reasoning chain — the skill graph traversal, the Bloom's level comparison, the alternatives it considered and rejected. Then it started remembering. After each session, it debriefs with the educator and extracts patterns. 18 sessions in, it knows hands-on exercises run 3 minutes over for evening cohorts. 37 MCP tools. 10 skills. 3 subagents. 9 moonshots. 24 frontend routes. 4 domains. Google Workspace integration.