Four Rules to Stop Your AI from Architecting Itself into a Corner

What it does This repo packages Andrej Karpathy’s specific complaints about how LLMs write software—silent assumptions, unchecked confusion, and gratuitous overengineering—into four concrete principles meant to constrain Claude Code. It ships as both a cross-project plugin and a per-project CLAUDE.md, plus a matching Cursor rule. The goal is to force the model to ask rather than guess, simplify rather than bloat, and touch only what you asked for.

The interesting bit The guidelines act as a behavioral patch for autocomplete’s worst instincts. “Surgical Changes” explicitly forbids “improving” adjacent code or matching your preferred style over the existing one, while “Goal-Driven Execution” demands verifiable success criteria instead of vague imperatives like “fix the bug.”

Key highlights

• Four rules: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution
• Forces explicit reasoning: state assumptions, present tradeoffs, and stop when confused
• Ships with a Cursor rule file so the same constraints apply in .cursor/rules/
• Designed to merge with existing CLAUDE.md files rather than replace them
• Explicitly admits it is biased toward caution over speed

Caveats

• The README’s installation instructions are vague; “Open fastsetup application” is not a universally recognized step
• At its core, this is a well-structured markdown prompt packaged as a plugin

Verdict Worth a look if Claude Code has ever silently rewritten your imports or architected a 1,000-line solution to a 100-line problem. If you already enforce strict prompting discipline manually, this formalizes your existing workflow without adding much else.