AI Agent Guidelines for CS336 at Stanford
Stanford's CS336 course provides guidelines for AI agents to act as teaching assistants, not solution generators, in programming assignments.
The document establishes rules for AI coding assistants (ChatGPT, Claude, GitHub Copilot, Cursor) helping students in CS336, an implementation-heavy machine learning course. AI agents should explain concepts, point to course materials and documentation, review student code, and debug through guiding questions. They must not write code or pseudocode, complete TODOs, implement core components like tokenizers or training loops, or provide direct solutions. The teaching approach emphasizes asking clarifying questions, referencing lecture materials, suggesting next steps rather than implementing them, and using tests and assertions over direct fixes. An example shows an agent directing a student to check mask application timing and broadcast behavior via a toy sequence example rather than identifying the bug.
What HN community is saying
Top comments dismiss the guidelines as unenforceable given AI's capabilities, arguing universities should instead increase assignment difficulty, use in-person exams, and redesign coursework around labs and large projects. One counterargument notes the approach is better than outright bans and acknowledges students will use AI regardless, so modeling healthy usage has value. Skeptics point out passive observation through an AI tutor may be worse than no help, noting students cheat themselves academically but face concrete GPA pressure. A thread notes the repo includes both CLAUDE.md and AGENTS.md duplicates, wishing Claude Code would check for both filenames.