A ten-rule CLAUDE.md file attributed to Andrej Karpathy is circulating in the developer community. Compared to the four-rule community version that went viral in January, the biggest change is not new content but a shift in gaze — from "how to write code" to "how to supervise itself." The first four rules cover the most common failure modes — think first, keep it simple, surgical changes, goal-driven. Among the six added rules, Verification forces "write the reproduction test before changing code," and Common Failure Modes calls out four self-recognized loss-of-control patterns by name: Kitchen Sink (fixing the faucet turns into remodeling the kitchen), Wrong Abstraction, Optimistic Path, and Runaway Refactor.
Worth noting: these rules are designed specifically for loops, because Anthropic's /goal command in Claude Code 2.1.139 has already architecturally separated "the model that writes code" from "the model that verifies code" — the goal is "knowing when to stop," not "finishing the task." The real meaning of the ten rules: an agent without per-turn human review must take on the reviewer's responsibility itself. Once that line is crossed, the first four rules are only the entry point; the latter six are the real engineering discipline.