Karpathy's antidote for AI assistants that overcomplicate by default

What it does This repository packages Andrej Karpathy’s observations about LLM coding failures into a single behavioral guide for AI assistants. Delivered as a CLAUDE.md file and a Cursor rule, it instructs coding agents to state assumptions explicitly, reject unnecessary abstractions, touch only relevant code, and verify work through tests rather than vague imperatives.

The interesting bit The project treats the system prompt itself as the product. Rather than fixing generated code after the fact, it attempts to pre-empt bad habits—overcomplication, drive-by refactoring, and silent confusion—by hardwiring four constraints into the agent’s context window.

Key highlights

• Four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution
• Directly targets Karpathy’s documented pitfalls: hidden assumptions, bloated abstractions, orthogonal edits, and deleted comments
• Ships as both a Claude Code plugin and a per-project CLAUDE.md file, with a committed Cursor rule for cross-editor use
• Designed to layer under project-specific instructions rather than replace them

Caveats

• The guidelines deliberately trade speed for caution; the README concedes trivial fixes don’t need the full rigor
• It is generic by design and requires merging with project-specific rules to handle domain constraints

Verdict Worth a look if your AI assistant keeps “improving” adjacent code or delivering 1,000-line solutions to 100-line problems. Less necessary if your existing prompts already enforce surgical minimalism.