A Karpathy-inspired CLAUDE.md Breaks 15K Stars: How a Markdown File Can Tame an AI’s Bad Habits of Writing Code

A GitHub project called andrej-karpathy-skills, which contains only a single Markdown file, breaks 15,000 stars and becomes one of the most popular open-source projects in the Claude Code ecosystem. This CLAUDE.md file is based on former Tesla AI chief Andrej Karpathy’s observations about common mistakes made when writing code with LLMs, turning them into behavior guidelines that can be used directly with Claude Code.

Common LLM programming pitfalls, as observed by Karpathy

Karpathy points out that when LLMs write code, they make some predictable mistakes: over-engineering, ignoring existing code patterns, and adding dependencies where they’re unnecessary. These aren’t random errors—they’re systematic biases caused by how the models are trained. The model tends to present “clever” solutions rather than concise ones that fit the project context.

The key insight is this: if these mistakes are predictable, you can prevent them with the right instructions. This is the practical application of “feedforward” in Harness Engineering—set the rules before the AI acts, rather than trying to fix things afterward.

How a single Markdown file can change AI behavior

CLAUDE.md is Claude Code’s project-level configuration file. When you place it in your project’s root directory, Claude Code automatically reads it and follows the instructions it contains every time it starts up. This file turns Karpathy’s observations into four core principles:

Goal-driven execution — convert imperative instructions into declarative goals, paired with a validation loop

Don’t assume — when you’re unsure, you must confirm first rather than guess

Don’t hide confusion — if you don’t understand the requirements, you must state it clearly

Actively expose trade-offs — when multiple options exist, present their respective pros and cons

These principles may sound like advice for human engineers, but in the context of AI they mean something different. The default behavior of LLMs is to “produce a complete response as much as possible,” even if that means guessing the user’s intent or over-designing. CLAUDE.md steers these default behaviors in a more cautious direction.

The trend behind the 15K stars: a new form of Prompt Engineering

The project’s explosive popularity reflects a shift in the developer community: evolving from “using AI to write code” to “the behavior of engineering with AI makes code quality better.” In the past, prompt engineering focused on crafting prompts for a single conversation; now the focus is on persistent behavior guidelines—set once, effective long term.

It also echoes an aspect of the Vibe Coding trend that hasn’t been discussed enough: when 92% of U.S. developers are already using AI programming tools, determining code quality is no longer just about model capability, but about how you “manage” the behavior of this AI teammate. A good CLAUDE.md may be more effective than choosing a stronger model.

The project was created by developer forrestchang, is 100% open-source, and—besides the main CLAUDE.md file—also provides versions that can be installed and used as Claude Code Skills.

This article, Karpathy-inspired CLAUDE.md breaks 15K stars: how a single Markdown file tames AI’s bad coding habits, first appeared on ChainNews ABMedia.