AI Agent Guidelines for CS336 at Stanford

Original Article ↗ Hacker News Discussion ↗

Stanford’s CS336 course is experimenting with written “AI agent” guidelines that tell tools like Claude how they may assist students—coaching, explaining, and reviewing code rather than generating or running it outright. Commenters are split between seeing this as a realistic middle ground that teaches healthy AI use in line with honor codes, and criticizing it as unenforceable, either too restrictive for real-world preparation or too weak to prevent students from offloading their learning to automated tools.

Overall reaction to the AI agent guidelines

  • Many see the guidelines as a reasonable, realistic middle ground between banning AI and letting it write all the code.
  • Others consider them “good intention but useless” because they rely on student self-restraint and are easy to bypass.
  • Some think elite institutions should go much further in rethinking curricula around AI, not just adding a README.

Enforceability and honor-code debates

  • Repeated concern: guidelines are fundamentally unenforceable; students can just use external models or edit CLAUDE.md / AGENTS.md.
  • Defenders argue enforcement is secondary; the value is in clearly stating “healthy use” norms and trusting students’ integrity.
  • There’s disagreement about how well honor codes work in practice; some claim they worked surprisingly well, others say cheating is common and largely invisible.

Learning vs. shortcutting

  • Many worry that easy access to AI encourages “cognitive laziness,” analogous to junk food vs. exercise.
  • Others argue students should be allowed full use of AI, with responsibility on instructors to design assessments that still test real understanding.
  • Several note that students can deceive themselves into thinking they’re learning when they’re passively watching AI or videos.
  • Some report direct experience: using AI heavily for code feels like “cheating myself” and harms skill-building.

Assessment design in an AI world

  • Strong support for high-stakes in-person exams (written, oral, or laptop-without-internet) to ensure students can perform without agents.
  • Suggestions include:
    • Harder, more conceptual assignments where agents struggle or can’t be blindly trusted.
    • Oral exams / code walkthroughs that quickly expose AI-generated work without understanding.
    • Weighting exams heavily and treating homework more as practice, even if some cheat there.

Use of AGENTS.md / CLAUDE.md and tooling

  • Discussion of using AGENTS.md / CLAUDE.md as a standard contract for how agents should behave in a repo.
  • Some think the Stanford version is too verbose and may fall out of context; others say long prompts are common and effective.
  • A few instructors are experimenting with similar files plus AI-usage histories to coach, not punish, overreliance.

Student culture and future skills

  • Reports that many teens both use AI and culturally “hate” it; knowing material without AI is seen as a social “flex.”
  • Employers in the thread split between:
    • Wanting students trained to use AI fully on hard problems.
    • Wanting deep fundamentals and general learning ability, not tool-specific optimization.