Inspiration
Students often have a Jupyter notebook that runs—or almost runs—but still cannot tell which concept they misunderstood. Traditional linters focus on code correctness, while general chat feedback often lacks cell-level evidence and a way to verify learning afterward.
Notebook Coach turns a course notebook into a short, evidence-linked coaching loop for students learning Python and large language model concepts.
What it does
Notebook Coach is a Codex Skill that:
- inspects a Python or LLM course notebook without executing it by default;
- identifies code errors, conceptual gaps, reproducibility issues, and unclear explanations with references to specific cells;
- generates a structured Markdown report and a separate challenge notebook containing exactly two targeted, answer-free tasks: one code task and one concept task;
- lets the learner improve the source notebook and complete the challenges;
- rechecks the work and produces a verification report showing resolved issues, remaining gaps, challenge status, and changes in a transparent Learning Evidence Score.
The score measures evidence visible in the notebook—not the learner's absolute ability. Every deduction is tied to a cell, severity, and versioned rule.
How we built it
GPT-5.6 Codex handles teaching judgment, concept diagnosis, targeted challenge design, and open-ended explanation review. A local Python 3.11 toolchain handles deterministic notebook parsing, secret redaction, bounded snapshots, risk scanning, hashing, scoring arithmetic, run state, artifact validation, and before/after comparison.
The architecture does not require a separate OpenAI API key or API billing path. It runs inside the learner's existing Codex session and uses local scripts for reproducible checks.
Safety and privacy
The source notebook is never rewritten. Static analysis is the default. Model-facing source, saved outputs, challenge answers, and runtime summaries pass through bounded redaction.
Optional execution is a separate two-phase action. Notebook Coach first shows the exact target, SHA-256 hash, kernel, temporary directory, and time limits, then executes only after explicit hash-bound confirmation. Execution uses a separately scanned temporary copy with per-cell and total timeouts; it is a safety boundary, not an operating-system sandbox.
What we accomplished
- Installable Codex Skill with a one-command installer.
- Evidence-linked diagnosis and exactly two answer-free challenges.
- Deterministic recheck and Learning Evidence Score.
- Safe Transformer attention sample plus complete pre-generated judge artifacts.
- 326 automated tests passing on Python 3.11.
- Fresh-checkout installation and acceptance verification.
- Public GitHub repository and a 1:58 public demo video.
Why it matters for education
Notebook Coach does more than fix code. It asks the learner to demonstrate understanding, keeps diagnosis separate from answers, and makes progress inspectable. Teachers and judges can review the report, challenge notebook, and verification artifact without a custom backend or database.
Links
Built With
- codex
- gpt-5.6
- jupyter-notebook
- nbclient
- nbformat
- python-3.11