Stanford's CS336 ships a CLAUDE.md telling AI tutors to coach, not code

Stanford’s CS336 has published a CLAUDE.md aimed at AI coding assistants — Claude Code, Copilot, Cursor, ChatGPT — that students point at the course repo. The instructions reframe the assistant as a teaching aid: explain concepts, review student-written code, interpret PyTorch/CUDA/Triton errors, and ask Socratic debugging questions, but refuse to write Python or pseudocode, complete TODOs, edit the student repo, run shell commands, or hand over implementations of tokenizers, transformer blocks, optimizers, training loops, kernels, or RL pipelines.

The document is unusually concrete about the failure mode it wants to prevent. Worked examples show the right response to prompts like “fix my tokenizer” or “tell me what my mistake is” — refuse the direct fix, ask what the student tried, and steer them toward shape assertions, toy inputs, and profiler data. Pointing students at third-party implementations is also off-limits, since the course is meant to be self-contained.

The interesting move is treating agent behavior as something a course can specify in-repo, the same way projects already use CLAUDE.md to constrain agents on production codebases. It’s an early example of educators trying to encode pedagogy into the agent’s system prompt rather than relying on honor-code language in a syllabus — and a bet that students will actually run assistants pointed at the repo instead of pasting prompts into a separate chat window.