The file, a community maintained distillation of tips from former OpenAI researcher Andrej Karpathy, codifies short behavioral rules that target how AI coding agents fail silently: filling in unstated requirements, over engineering simple tasks, and refactoring code unrelated to…
A plain-text rules file for AI coding agents has become one of GitHub's most-starred AI repositories in 2026, not because it unlocks new model capability, but because it imposes a discipline the models lack on their own.
The artifact is a roughly 65-line Markdown file called CLAUDE.md, sitting at the root of a public repository named [andrej-karpathy-skills](https://github.com/multica-ai/andrej-karpathy-skills/blob/main/CLAUDE.md). In April 2026, the repository passed 176,000 stars according to Chinese tech outlet Leiphone's reporting, with a corroborating X reply placing the count above 180,000 shortly afterward. In the same window, Anthropic's official skills repository sat at roughly 151,000 stars. A community-curated distillation of public tips from former OpenAI research lead Andrej Karpathy, not an official Karpathy or Anthropic artifact, outpaced the vendor it was nominally built for.
That inversion is the story. The file's argument is that AI coding's binding constraint has moved from raw capability to unsupervised behavior. Karpathy himself framed the shift in a public post: in November he was still running an 80/20 mix of manual coding with autocomplete and one-off agent calls. By the time of his post, that ratio had flipped to roughly 80% agent output and 20% targeted edits and touchups. Once the model is doing most of the writing, the question stops being whether it can write code and starts being whether it can stop itself from writing wrong code.
Three principles are visible in the fetched excerpt of the file. Leiphone's coverage refers to "four points," suggesting a fourth rule lives further down in the document that is not visible in the truncated public excerpt. The principles below are quoted or paraphrased directly from the visible portion.
Think before coding. The first rule treats ambiguous specifications as a feature rather than a nuisance. Rather than filling in unstated requirements silently, the agent is told to surface assumptions, ask before guessing, and present tradeoffs. The failure mode it targets is the most expensive one in unsupervised agent work: code that compiles, passes tests, and quietly solves the wrong problem.
Simplicity first. The second rule instructs the agent to default to the minimum viable implementation: no speculative features, no premature abstractions, no configurability added "just in case." The rule is explicit that a 200-line implementation that handles the stated need should be rewritten as 50 lines rather than left as-is. Leiphone highlights this rule as the file's signature idea, and cites an anecdotal developer workflow in which installing the rule lifted a task pass rate from 65% to 94%, a self-reported number with no published benchmark and no replication data, useful as illustration rather than as evidence.
Surgical changes. The third rule constrains the scope of any modification. The agent should touch only what the requirement forces it to touch; it should not reformat adjacent code, refactor unrelated modules, or delete comments it does not understand. The failure mode here is modification sprawl: the agent decides that while it is "in the area" it should also clean up, and the diff balloons, breaking review and reviewability.
A fourth principle is referenced in the Leiphone summary but not visible in the truncated file excerpt. The most plausible candidate, based on the surrounding principles and Karpathy's public commentary, is a verification or "done means done" rule: the agent should not declare a task complete without an objective, reproducible success check. Readers should treat that fourth principle as a working hypothesis to confirm against the full file before adopting it.
The pattern matters beyond Karpathy's brand. CLAUDE.md is one instance of a broader rules-file convention: short, version-controlled Markdown at the root of a repository, read by the agent before any code is written, edited by the developer as the team's working agreement with the model. Several vendors, Anthropic first, but also Cursor and others, now treat the file as a first-class surface for steering agent behavior. The popularity of andrej-karpathy-skills suggests developers are starting to share and remix those files the way they share .editorconfig or .eslintrc files.
What the file gets right, and what it does not. The rules address real, observable failure modes. A developer who reads the file and writes their own CLAUDE.md this week will probably ship fewer "helpful but wrong" pull requests next month. But the artifact is narrow in the ways narrow artifacts are: it carries no security review, no deployment guidance, no model selection advice, no evaluation rigor. The 65%-to-94% pass-rate lift is one developer's experiment, not an industry benchmark. And the repository's name is borrowed; Karpathy has not claimed it as his own, and Anthropic is not the publisher.
The honest read is that a 65-line file beat Anthropic on GitHub stars not because it contains secrets, but because it names a constraint developers were already feeling. The model is capable enough. The thing it cannot do on its own is stop.