andrej-karpathy-skills (multica-ai mirror) is a single CLAUDE.md file that encodes four behavioral rules derived from Andrej Karpathy's documented frustrations with LLM coding agents. The file is 70 lines, human-readable end-to-end, and not project-specific — it can be copied into the root of any repository to produce an immediate change in how Claude Code writes code. The repo has crossed 132,000 stars on the multica-ai org mirror plus another 91,000 on the original forrestchang account (220,000+ combined), held #1 on GitHub Weekly Trending for 28 consecutive days, and sits at position 94 in the global GitHub star ranking as of May 2026. The four rules target the most common Karpathy-cited LLM failure modes: scope creep, premature abstraction, comment hallucination, and silent backward-compatibility shims.
andrej-karpathy-skills is the most-starred Claude Code configuration file on GitHub. The premise is unusually simple: take Andrej Karpathy’s well-documented complaints about how LLMs write code, distill them into four behavioral rules, and put them in a CLAUDE.md file that any project can adopt with one cp command. No setup, no MCP server, no agent harness change — just a markdown file Claude Code reads.
The reception has been one of the most decisive GitHub trending events of 2026. The multica-ai mirror crossed 132,000 stars; the original by Forrest Chang sits at 91,000+; together they’ve held #1 on GitHub Weekly Trending for 28 consecutive days. As of May 23, the file is still gaining ~3,372 stars per day — sustained velocity that’s rare even for breakout projects.
The file does one thing well: shape agent behavior at the prompt-layer instead of through tool changes or harness modifications. The four rules target Karpathy’s most frequently cited LLM coding failure modes — silently expanding scope beyond what was asked, introducing premature abstractions for hypothetical future needs, writing comments that describe what code does rather than why, and adding backwards-compatibility shims when a clean rewrite was the right call. Because the rules are generic, they apply unchanged across web app, ML, embedded, and infrastructure codebases. The file is 70 lines and takes under a minute to read end-to-end before adopting — a property that distinguishes it from the larger “skill pack” frameworks that require harness setup.
andrej-karpathy-skills is a baseline addition for any new Claude Code project — the cost is one cp of a 70-line file, the benefit is immediate behavioral correction on the failure modes most users hit within their first few sessions. It pairs cleanly with project-specific CLAUDE.md content; you can append your repo-specific conventions below the Karpathy rules and Claude Code will apply both. It is also a popular starting point for teams writing their own skills packs, because the rules give a clear template for what “behavioral guidance” should look like at the markdown layer.
AI agents that work well with Andrej Karpathy Skills.
Production-grade engineering skills for AI coding agents — addyosmani's curated bundle covering Claude Code, Cursor, and Antigravity workflows.
100+ ready-to-use Claude Code agents, commands, hooks, and integrations — installable via CLI or interactive dashboard at aitmpl.com.
Injects up-to-date, version-specific documentation into LLM context — eliminates hallucination from stale training data.