Andrej Karpathy describes Claude Code agent workflow

AI researcher Andrej Karpathy builds complete software projects solely by conversing with Claude Code using short prompts, diagrams, and iterative subagent pipelines. Rather than writing or reading code directly, he relies on cheap filters to process large datasets before routing relevant portions to smaller models for bulk work and larger models for orchestration. When sharing his work, he distributes the underlying data and prompt rather than the application code. Open article →