A ten-rule version of the CLAUDE.md file attributed to Andrej Karpathy began circulating on X on Friday, according to TechTimes. The document adds six rules to the four-rule community template that has accumulated over 200,000 combined GitHub stars across two repositories since early 2026. Karpathy, who joined Anthropic’s pre-training team five weeks ago, has not publicly confirmed the document’s authenticity.
The original four rules address the most common LLM coding failure modes: think before coding (state assumptions explicitly), simplicity first (minimum viable code only), surgical changes (no edits outside the stated task), and goal-driven execution (convert vague instructions into verifiable outcomes). Developer Forrest Chang distilled these from Karpathy’s public observations into the andrej-karpathy-skills repository, which became one of the fastest-growing files in GitHub history.
Six New Rules for Post-Generation Discipline
The circulating document’s additions target a gap the original four rules did not cover: what happens after the code is written, when the agent is running, evaluating, and deciding whether to continue.
Verification establishes a specific order: before attempting to fix a bug, write a test that reproduces it reliably. Fix the code. Run the test. Only when the test passes is the bug resolved. In autonomous loop contexts where no human reviewer checks each step, the test becomes the only validation checkpoint.
Goal-Driven Execution in the expanded version goes further than the community file. The document requires defining what “done” looks like in machine-verifiable terms before any code is written. “Add validation” fails this standard. “Users who submit a blank or malformed email field see a specific error message, and both cases have passing tests” passes it. For multi-step work, a plan must come first.
Debugging gets its own rule with a prescribed sequence: read the full error and stack trace, reproduce the problem before attempting a fix, and change one variable at a time. The failure mode this addresses is confident wrong diagnosis, where an agent reads an ambiguous error, picks an interpretation, and generates a fix for a problem it has not confirmed exists.
Dependencies treats every added package as permanent third-party code updated on someone else’s schedule. Before reaching for a library, check whether the standard library handles it. If a dependency is added, document the decision explicitly.
Communication draws a line between useful uncertainty and vague reassurance. “I’m not sure this library supports streaming” is actionable. “I think this should work” is not. The rule prohibits confident-sounding guesses when uncertainty is the accurate answer.
Common Failure Modes names four recurring patterns explicitly, giving the agent a checklist of known traps to watch for during execution. TechTimes reports this is the most distinctive of the six additions, as it moves from process rules to pattern recognition.
From Prompt Engineering to Engineering Discipline
The original four-rule CLAUDE.md went viral because it addressed pain points that anyone using Claude Code for more than a few sessions had experienced: speculative abstractions, unrequested features, edits to code outside the task scope. The six additions shift the document’s scope from prompt engineering (how to instruct the agent) toward engineering discipline (how the agent monitors its own execution quality).
The timing coincides with Karpathy’s move to Anthropic’s pre-training team in May. Whether the expanded document reflects internal Anthropic practices, Karpathy’s personal coding workflow, or a community evolution remains unconfirmed. What developers on X responded to, according to TechTimes, was the content itself: rules that address the gap between code that seems correct and code that actually runs in production.