Updated
Updated · Tech Times · Jun 28
Karpathy-Attributed 10-Rule CLAUDE.md Adds 6 Self-Check Rules for AI Coding Loops
Updated
Updated · Tech Times · Jun 28

Karpathy-Attributed 10-Rule CLAUDE.md Adds 6 Self-Check Rules for AI Coding Loops

3 articles · Updated · Tech Times · Jun 28

Summary

  • A 10-rule CLAUDE.md file attributed to Andrej Karpathy spread on X on Friday, extending the widely used 4-rule template with six additions aimed at autonomous coding loops rather than one-off prompting.
  • Those six rules focus on self-monitoring after code is written: verify with reproducible tests, define machine-checkable goals, debug step by step, limit dependencies, communicate uncertainty clearly, and stop on four named failure patterns.
  • Developers say the shift matters because loop-based workflows let agents run, evaluate and continue without human review; Anthropic's Boris Cherny recently described his role as writing loops rather than prompts.
  • The file's authenticity remains unconfirmed, and Karpathy has not commented. CLAUDE.md is injected as project context at session start, influencing behavior but not enforcing it and leaving room for prompt injection or malicious files.
  • The circulating document builds on a community repository distilled from Karpathy's January posts that has topped 200,000 combined GitHub stars, underscoring growing demand for stricter controls on agentic coding costs and errors.

Insights

As new AI rules risk code quality for efficiency gains, how can developers find the perfect balance for different models?
Are these rules teaching AI to genuinely reason, or just creating more disciplined instruction-following machines?
As AI agents become autonomous online actors, what new frameworks can guarantee they are both secure and accountable?