Stanford's CS336 Course Requires Students to Disclose AI Use as a 'Partner' in Assignments

Stanford University's CS336 Deep Learning Systems course has published a CLAUDE.md configuration file in its GitHub repository that sets explicit rules for AI coding agents like Claude Code, ChatGPT, GitHub Copilot, and Cursor when used on course assignments. The file instructs these tools to act strictly as teaching aids rather than solution generators.

Defining the Role of AI in Deep Learning Education

According to the repository file, CS336 is designed to be intentionally implementation-heavy with limited scaffolding, requiring students to build foundational deep learning infrastructure from scratch. To preserve that learning process, AI coding assistants are instructed to function as interactive teaching aids that guide students through dialogue rather than write code for them.

AI agents are explicitly instructed not to write any Python or pseudocode, edit student repositories, or complete TODO sections. Instead, agents are tasked with reviewing student-written code, explaining error messages, and directing students toward course materials at cs336.stanford.edu.

Prohibited AI Capabilities and Task Boundaries

The file lists clear boundaries for what AI agents may not do:

• No direct code generation: Agents cannot write Python code, generate pseudocode, or refactor student implementations into finished scripts.
• No automatic implementation: Tools are barred from completing core assignment components, including tokenizers, transformer blocks, optimizers, training loops, Triton kernels, distributed training logic, scaling-law pipelines, data filtering pipelines, and RL alignment methods.
• No environment control: Assistants must not execute bash commands or directly edit files inside the student's project repository.
• No third-party shortcuts: Agents are forbidden from directing students to external third-party implementations, keeping the coursework self-contained.

Guiding Debugging Without Giving Away Answers

When students encounter bugs or performance bottlenecks, the file directs AI agents to respond with diagnostic questions rather than fixes — for example, suggesting sanity checks, shape assertions, or profiler investigations for CUDA, PyTorch tensor, or Triton issues.

Example prompts in the file show an assistant facing a failing causal mask issue asking the student to check tensor broadcasting, masking order relative to softmax, and negative values. For performance issues in a Byte Pair Encoding (BPE) tokenizer or distributed training setup, the AI is told to guide students to separate compute from communication time rather than deliver optimized code.

Limitations and Enforcement Challenges

Because these instructions live in a repository file that depends on agent compliance, students could bypass the constraints by using unconstrained external AI tools. The framework therefore relies on student cooperation rather than technical enforcement.

Why it matters

The CS336 file is a concrete example of how a computer science course is trying to integrate AI coding assistants without letting them substitute for the hands-on implementation work the course is built around — a template other instructors may look to as AI tools become more capable.