Inspiration
ARC-AGI-3 puzzles are a great way to practice observation, hypothesis-building, and reasoning, but teenagers exploring them alone often get stuck without any support. We wanted a coach that helps them think through problems without ever handing them the answer.
What it does
TeenARC lets teens play the official public ARC-AGI-3 games while a Socratic AI coach nudges them to observe changes, form testable hypotheses, and reflect on evidence. It renders every game frame, tracks an evidence timeline, resets, levels, and action efficiency, and never reveals rules, routes, or solutions.
How we built it
The frontend is React and TypeScript; the backend is FastAPI wrapping the official arc-agi Python toolkit. Coaching prompts are routed through the local Codex CLI, which only ever receives booleans and numeric session signals, never learner text or game solutions, and must pick from a small set of reviewed question templates.
Challenges we ran into
Keeping the coach genuinely Socratic (never leaking a solution) while still being useful required a strict schema-constrained interface between the app and the LLM, plus a final no-solution gate before any question reaches the learner.
Accomplishments that we're proud of
We built a working Socratic coach layer where the LLM never sees learner text, game frames, or solutions, only bounded booleans and numbers, yet it can still guide a teen through observation and hypothesis testing. We also shipped replay controls, an evidence timeline, and accessibility features like reduced motion and keyboard coordinate entry in the same build.
What we learned
Constraining an LLM to a small set of vetted actions can make an AI tutor much safer for teenage users without making it less helpful.
What's next for TeenARC
Expanding the template library, adding more accessibility support, and running structured pilots with educators.
Built With
- arc-agi
- cli
- codex
- fastapi
- python
- react
- typescript
Log in or sign up for Devpost to join the conversation.