Stop Prompting Claude — Use Karpathy’s Method Instead
- Video: Stop Prompting Claude. Use Karpathy’s Method Instead. by Austin Marchese (13:18)
- Tags:
claudekarpathyspec-drivenworkflow
Key Points
Breaking down how Andrej Karpathy (ex-Tesla AI head, speaking at AISN 2026) builds ~10x faster — three layers:
- The limitation: AI is brilliant at what can be measured, but blind to context. Karpathy’s example: “car wash 50m away — drive or walk?” Every frontier model says walk; they miss that you need the car at the wash. (Confirmed across Claude, Gemini, Grok, ChatGPT.)
- Layer 1 — The Spec. How you deliver your understanding to Claude in a usable format. Plan mode is “too high-level” — Karpathy wants a detailed spec co-designed with the agent.
- Uncover the goal, not the task. “Create an end-of-month report” is a task; the goal is the decision it drives — something AI can never decide for you. Tell Claude to interview you to extract it.
- Layer 2 — Be agile, not waterfall. Don’t have the agent build the whole thing then reveal it; break into small buckets and review results throughout.
- Layer 3 — the one thing to focus on in the age of AI (the spec/taste/judgment layer).
Related
- karpathy-second-brain-ai-os — “I Turned Karpathy’s Second Brain Into an AI Operating System” (prev video)
- llm-wiki — Karpathy’s LLM Wiki concept