The linked file is a set of instructions for coding agents used in Stanford's CS336 course. It tells the model to help students learn by explaining concepts, asking questions, and guiding debugging, while avoiding actions that would short-circuit the assignment such as directly producing the full solution or running commands on the student's behalf. People liked the basic posture because it accepts that students will use AI anyway and tries to channel that toward tutoring instead of pretending a ban will hold. The consensus was that this is useful as a norm-setting layer, especially when shipped as AGENTS.md or CLAUDE.md inside the assignment repo so tools pick it up automatically.
Nobody thought the file itself solves the cheating problem. Students can edit or ignore it, use a different tool, or just route around it entirely. The useful reading is narrower and more practical. This gives well-meaning students a better default experience, and it gives instructors a clearer statement of what healthy AI use looks like. Most of the serious comments landed on the same point. Learning has to be validated somewhere the student cannot outsource understanding. That means oral walkthroughs, in-person exams, heavily weighted tests, or other assessments that force explanation and transfer, not just polished submitted code.
A second theme was that agent behavior is shaped less by a giant wall of instructions than by environment and tooling. Several people using
Claude Code said long policy files get ignored or drift out of the active context, and that if a course really needs things like prompt logs or constrained behavior, hooks and
harness-level controls are more reliable than asking the model nicely. That pushed the conversation from prompt-writing toward system design. If you care about compliance, make the tool do it.
The thread was also unusually blunt about the education tradeoff. Many commenters think unrestricted AI use lets students move faster while quietly hollowing out the struggle that makes concepts stick. A physics professor reported a sharp drop in help-session attendance and worse outcomes among weaker students, which matched the broader fear that AI tutoring can feel like studying while keeping the student in a passive role. The thread did not end up anti-AI. It ended up pro-scaffolding. Use agents, but design the course so students still have to think, explain, and own the result.
github.com