Karpathy Skills
Andrej Karpathy-inspired CLAUDE.md guidelines that make AI coding agents less chaotic
Karpathy Skills is a single-file GitHub repository containing a CLAUDE.md that encodes four principles Andrej Karpathy identified as the core failure modes of LLMs writing code: silent assumptions, overengineering, unnecessary orthogonal changes, and vague goal-setting. The file is designed to be dropped directly into any project and works as a Claude Code plugin or a plain CLAUDE.md, shaping the model's behavior before it starts modifying your codebase. The four principles are: Think Before Coding (force explicit reasoning and ask clarifying questions before touching anything), Simplicity First (the minimal working solution is always the right default), Surgical Changes (only edit what the task requires — nothing else), and Goal-Driven Execution (transform every vague request into a set of verifiable success criteria before starting). The repo went from zero to 9,200 stars in its first weeks, making it one of the fastest-growing 'prompt engineering for agents' resources ever published. What's interesting about the project is what it reveals about the current state of AI coding agents. Despite massive capability improvements, the failure modes that Karpathy documented in early 2026 — models that silently assume, over-reach, and over-build — remain stubbornly present. Karpathy Skills is a behavioral patch applied at the project level: cheap, portable, and immediately testable. For teams running Claude Code or any CLAUDE.md-compatible agent, it's a five-minute install with measurable impact on code review overhead.
Panel Reviews
“”
“”
“”
“”
Community Sentiment
“”
“”
“”