Inspiration
AI tutors are excellent at producing answers, but a polished explanation does not guarantee that a student's mental model has changed. A learner who made one hidden assumption often receives a wall of correct text without ever seeing the exact moment their reasoning diverged. We wanted to make the student's own attempt—not the model's answer—the center of the learning experience.
What it does
A student invokes AhaLoop Coach in Codex, pastes a problem and their reasoning, or attaches a photograph of handwritten work. The plugin identifies the earliest unsupported inference, states the hidden belief behind it, and asks the student to make a prediction. Crucially, guided mode stops here.
After the learner commits, GPT-5.6 creates the smallest safe counterexample that distinguishes the learner's belief from the correct rule. AhaLoop asks the learner to explain the observation before giving a concise explanation, then finishes with a transfer problem using different surface details.
The companion web prototype demonstrates the same attempt → diagnosis → prediction → counterexample → transfer interaction in a visual interface and runs locally without an API key.
How we built it
The primary project is a repository-distributed Codex plugin containing a validated plugin manifest, local marketplace entry, and an ahaloop-coach Skill. The Skill acts as a four-stage pedagogical state machine and includes counterexample guidance for algebra, geometry, probability, physics, logic, and beginner programming.
GPT-5.6 runs directly inside Codex and uses the learner's visible work as context. No Platform API key, hosted service, MCP server, or third-party account is required. The plugin never executes learner-provided code and marks ambiguity rather than inventing missing reasoning.
The responsive web prototype uses a dependency-free Node.js server and HTML/CSS/JavaScript frontend. Its curated demonstration remains fully testable offline.
How we used Codex
Codex was used throughout the build: interpreting the hackathon requirements, narrowing the concept, designing the learning loop, implementing the interface, creating the structured diagnosis prototype, reviewing safety tradeoffs, packaging the native plugin, writing judge scenarios, validating installation metadata, testing the Skill in a fresh task, and preparing this submission.
When we learned that the hackathon grant was for Codex credits rather than Platform API credits, Codex helped re-architect AhaLoop from an API-dependent web app into a native Codex plugin that uses GPT-5.6 directly. That made the judge path accessible without separate billing while strengthening the role of Codex in the product.
Challenges
The hardest product question was how to make an AI diagnosis actionable without turning the experience into another answer-generating chat. AhaLoop solves this with enforced pauses: locate one first break, ask for a falsifiable prediction, wait, show one minimal counterexample, and only then explain.
The second challenge was making the project genuinely testable without API credits. Packaging the pedagogy as a local Codex plugin eliminated credentials and backend setup while preserving multimodal GPT-5.6 reasoning.
Accomplishments
- Complete attempt → diagnosis → prediction → experiment → transfer loop
- Native, installable Codex plugin powered directly by GPT-5.6
- Typed and photographed student-work support through Codex context
- No API key, API balance, hosted backend, or third-party service required
- Responsive visual prototype rather than a prompt-only concept
- Reproducible judge scenarios for algebra, probability, and programming
What we learned
The most useful unit of tutoring is often not an explanation but a falsifiable belief. Once a student's hidden assumption becomes something they can predict and test, the correction becomes memorable. We also learned that a carefully designed Skill can turn a general coding agent into a disciplined tutor while keeping installation dramatically simpler than an API application.
What's next
Next steps include teacher-facing misconception summaries, more visual counterexample formats, accessibility improvements, and an evaluation set comparing first-break diagnoses against human tutor annotations.
Built With
- codex
- codex-plugins
- codex-skills
- css
- gpt-5.6
- html
- javascript
- node.js