🧠 LLM Usage Rules by Phase

For Neural Network Architecture Mastery Curriculum

(One section for each phase in both Version 1 and Version 2)

🔰 Phase 0 — Training Stack Ownership (V1)

Goal: Build a pipeline from scratch.

LLM Policy:

• ❌ No code generation or boilerplate help
• ❌ No asking for training loops, dataloaders, model class templates
• ✅ Only allowed: post-hoc error explanation (after you’ve tried 5+ min yourself)
• ✅ Optional: compare your final solution after finishing

Rationale: Phase 0 is a test of raw autonomy. LLMs here = crutch.

🧱 Phase 1 — Architectural Breadth (V1)

Goal: Modify existing architectures and understand common components.

LLM Policy:

• ✅ Ask for explanations of blocks (e.g., why BatchNorm helps)
• ✅ Ask for critiques of your own model variants
• ❌ No asking "what changes should I make to improve X?"