Karpathy CLAUDE.md Grows to Ten Rules: New Self-Check Protocol for AI Coding Loops

Key Points:

• A ten-rule document attributed to Andrej Karpathy, recently joined to Anthropic’s pre-training team, extends the widely used four-rule community template for AI coding agents by adding six new rules focused on agent self-monitoring and loop-level behavior, emphasizing verification, debugging, dependency management, communication, and recognizing common failure modes.
• The original four rules address common coding failure modes during code generation, while the six new rules target autonomous loop execution where the agent runs and evaluates code without human oversight, ensuring the agent verifies fixes with tests, defines clear goals, and stops when encountering recognized failure patterns.
• The document’s guidelines are injected into Claude Code sessions as behavioral context via a CLAUDE.md file, influencing but not enforcing agent behavior; this mechanism allows the agent to read the rules alongside system prompts but can be overridden by specific prompts or malicious files, underscoring the importance of obtaining the file from trusted sources.
• This development reflects a broader shift in AI-assisted development from manual prompt engineering toward designing autonomous loops that prompt, evaluate, and correct AI-generated code iteratively until verifiable goals are met, a practice termed "loop engineering" by Anthropic’s Claude Code creator.
• Although the document’s authenticity is unconfirmed and Karpathy has not commented, developers report that applying these ten rules changes the agent’s behavior significantly, fostering more cautious, test-driven, and goal-oriented coding workflows, which also help control the high token costs and risks associated with autonomous AI coding loops.