Karpathy's Internal Claude.md Leaked: Ten Commandments and the Loop Engineering Paradigm

Five weeks after former OpenAI researcher Andrej Karpathy joined Anthropic, his team's internal Claude.md configuration file was accidentally leaked. The file contains ten behavioral rules designed to guide the AI coding assistant in generating code that requires no rewrites. Compared to the community version (4 rules) with 180,000 stars on GitHub, the internal version adds six new rules covering verification, goal-driven execution, debugging, dependency management, communication, and common failure modes, emphasizing AI self-checking and engineering collaboration. Meanwhile, Claude Code founder Boris Cherny proposed the concept of "Loop Engineering," where automated loops enable AI to continuously execute tasks and self-correct, shifting the human role from operator to system designer. Karpathy's rules provide self-checking standards for the loop, marking a paradigm shift in AI programming from prompt engineering to loop engineering.