Karpathy ranted. Someone shipped it. Your AI behaves now.

Your AI keeps rewriting code you told it not to touch, then swears it's sure when it's dead wrong. There's a 65-line text file that fixes both — and it became one of the most-starred files on GitHub. Here's the part nobody mentions: Karpathy didn't even write it.

The origin

Andrej Karpathy posted four complaints about AI coding on Twitter. Silent assumptions. Bloated code. Collateral damage to files nobody asked about. Fake confidence.

A developer turned those four rants into four rules, dropped them in a file, and open-sourced it. No secret model. No magic prompt. Just house rules — the kind you'd give a good junior dev.

The system

Four rules. That's the whole thing.

Ask first. Don't assume.

Official · Think Before Coding

If anything is unclear, it asks before writing a single line. Assumptions get stated out loud; two readings of a request get presented, not silently picked.

kills most of the rubbish on its own

The simplest thing that works.

Official · Simplicity First

Minimum code that solves the stated problem. No 200-line "solution" when three lines do the job, no speculative abstractions.

3 lines. done.

Don't touch unrelated code.

Official · Surgical Changes

If a file wasn't part of the task, it doesn't get to "improve" it. No drive-by refactors, no reformatting, no renames in working code.

stops it breaking what already worked

Say when you're not sure.

Official · Goal-Driven Execution

If it's guessing, it tells you — instead of handing you something confidently broken. Vague asks become checkable success criteria.

honest beats confident

How to install

Drop it in once.

Drop the rules in a file called CLAUDE.md at the root of your project. Claude Code reads it automatically, every session. Not just Claude — the same four rules work as a rules file in Cursor and Codex.

~/your-project/CLAUDE.md

# House Rules

## 1 · Think before coding
State assumptions explicitly. If uncertain, ask.
If multiple interpretations exist, present them — don't pick silently.

## 2 · Simplicity first
Minimum code that solves the problem. Nothing speculative.
No features beyond what was asked.

## 3 · Surgical changes
Touch only what the task requires.
Unrelated code, comments, formatting: leave them alone.

## 4 · Goal-driven execution
Convert vague asks into verifiable success criteria.
If you're guessing, say so — before coding, not after.

Reality check

What actually changes.

✕ Before

Ask for a button change, get a refactored auth flow. 200 lines you didn't want, two files you didn't mention, and an AI that's "sure" right up until production breaks.

✓ After

It asks the one clarifying question first, ships the three lines that solve it, leaves everything else untouched, and flags exactly where it's unsure. Like a junior dev you'd actually promote.

Resources