Stanford released a public guideline for AI coding assistants used in its CS336 graduate class. The document spells out that these tools—ChatGPT, Claude Code, GitHub Copilot, Cursor, and similar—must act as teaching aides, not solution generators. Students are expected to write substantial Python and PyTorch code, so the guidance stresses preserving the hands‑on learning experience. The rulebook is hosted on the course GitHub.

The guidelines list concrete do’s: explain concepts, point to lecture notes on cs336.stanford.edu, review student code for edge cases, and ask probing questions instead of supplying fixes. Agents may clarify Python, PyTorch, CUDA, Triton, or distributed‑training errors, suggest sanity checks, toy inputs, or profiler‑based investigations, and keep dialogue focused on understanding rather than code delivery. These practices align with broader university AI use policies.

Prohibited actions include writing any Python or pseudocode, completing TODO sections, editing the repository, or generating full tokenizers, transformer blocks, or training loops. When a request crosses these lines, the agent must refuse and pivot to explanation or refer the student to staff. The policy aims to uphold academic integrity while allowing low‑level programming help within a self‑contained course. Compliance will be monitored through code reviews and staff oversight.