karpathy-skills

Overview

karpathy-skills is the most-starred CLAUDE.md template on GitHub — a single configuration file derived from Andrej Karpathy’s analysis of how large language models predictably fail when writing code. Forrestchang distilled Karpathy’s observations into four behavioral principles that correct for the most common Claude Code failure modes: making silent assumptions, overcomplicating solutions, scope creeping into adjacent code, and declaring victory before properly verifying output.

The repo hit #2 on GitHub Trending globally on April 13, 2026 with 5,828 stars in a single day — the fastest-growing CLAUDE.md template ever. The virality is recognition, not discovery: developers who’ve used Claude Code for more than a week have hit every failure mode the file addresses.

The Four Principles

Think Before Coding. Before implementing anything non-trivial, Claude must state its assumptions explicitly. If there are multiple valid interpretations, present them. If something is unclear, halt and ask — don’t fill in the gaps with whatever seems plausible and charge ahead.

Simplicity First. Implement the minimum code that solves the problem. No unrequested features, no abstractions for single-use code, no error handling for impossible scenarios. The self-test: “Would an experienced engineer view this as overengineered?”

Surgical Changes. When modifying a file, don’t “enhance” surrounding code, reformat things that weren’t broken, or refactor patterns that work. There’s a sharp distinction between dead code you introduced (clean it up) and pre-existing dead code (flag it, don’t touch it).

Goal-Driven Execution. Convert every task into measurable objectives with explicit verification steps before starting. “Add a login form” becomes “add a login form — success when: renders at /login, handles valid credentials, shows error on invalid, passes existing auth tests.”

Installation

Install as a Claude Code plugin (applies across all projects):

claude plugins install forrestchang/andrej-karpathy-skills

Or drop it directly into a project:

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

Use Cases

Solo developers hitting repetitive LLM behavior issues (over-scoping, assumption-making, noisy diffs).

Teams that want a shared behavioral baseline for Claude Code across projects — merge the four principles with project-specific CLAUDE.md content.

Anyone transitioning from manual coding to agent-driven development who needs a framework for correcting the behavioral regressions that come with heavier LLM use.