Stop Prompting Claude. Use Karpathy's Method Instead. - YouTube by Austin Marchese
- Move Beyond Vibe Coding / Random Prompting: Relying on spontaneous, unstructured prompts leads to drifting code quality and context limits. Instead, treat AI agents like junior developers or interns that require strict guidance.
- The Importance of Specifications (Specs): Before letting an AI touch code or write content, start with a rigorous planning phase. Define the core goals, requirements, and user decisions upfront.
- Agile, Step-by-Step Buckets: Break large projects down into small, manageable components. Build one bucket at a time using a strict cycle:
- Plan: Have the AI propose a detailed implementation plan for approval.
- Checkpoint: Review the output at strict milestones before moving forward.
- Verify: Manually test and verify edge cases (as AI can miss logic gaps, such as unique user ID mappings or hidden edge cases).
- Verifiability over Specification: As Karpathy notes, AI excels at tasks where outputs can be automatically checked (like code or math). Humans must stay in the loop to handle the "jaggedness" of AI—catching logical flaws that models miss (e.g., the famous AI car wash logic test).