assignment1-basics/CLAUDE.md at main

Takeaways

• Stanford CS336 (Language Modeling From Scratch) CLAUDE.md: AI agents are TAs, not solution generators for implementation-heavy assignments.
• Strict prohibitions: no Python/pseudocode, no solutions, no TODO completion, no repo edits, no bash commands, no third-party implementation pointers.
• Allowed: concept explanation, lecture/doc pointers, general code review feedback, guiding debug questions, high-level algorithm nudges, sanity checks/toy examples/profiler hints via dialog.
• Teaching loop emphasizes clarifying questions, lecture references, next-step suggestions, and tests/invariants over fixes (masks, shapes, toy tensors).
• Covers core assignment surface: tokenizers, transformers, optimizers, training loops, Triton, distributed training, scaling laws, alignment/RL—agents must not implement these for students.

Notes

Stricter than 1cg’s sample AGENTS.md gist (which allows 2–5 line examples). Reference for maximum academic integrity guardrails vs. vault productivity agents that actively implement.

Open questions

• Useful negative template when defining what vault agents should do that course agents must not?