First, What This Actually Is
It's a text file. Called CLAUDE.md. You drop it inside your project folder. Claude reads it before doing any work and behaves differently because of what it says.
That's genuinely the whole trick. No installer. No app. No subscription. Just a file with instructions Claude follows.
What's in those 65 lines? Rules. Four of them. Each one fixes a specific bad habit AI has when it writes code — making silent assumptions, overcomplicating simple things, rewriting code you didn't ask it to touch, and shipping work without checking if it actually does the job.
This Isn't Some Random Dev's Hot Take.
Andrej Karpathy's resume in one line: co-founded OpenAI, led Tesla's self-driving AI, taught Stanford's deep learning course. When he tweets about how AI codes, people listen. When he releases a file, it hits #1 GitHub trending by lunchtime.
The file started as a long X post where he was venting about how LLMs "make wrong assumptions on your behalf and just run along with them." A developer named Forrest Chang turned that venting into the four rules in this file. Multica AI mirrored it. Now everyone's using it.
65% to 94% — But Read The Fine Print.
You'll see this number everywhere: 65% to 94% accuracy.It's real. But here's what it actually measures.
The 94% is how often Claude follows the four rules in the file. It's not a measure of bug-free code. The file doesn't make Claude smarter. It makes Claude more disciplined. Fewer silent decisions. Fewer surprise rewrites. Fewer "why did you also touch this other file?" moments.
That's still huge. Most of the pain of working with AI on code isn't the model's intelligence — it's the model's habits. Fix the habits and the whole experience changes.
Two Paths. Pick One.
Path A · If you use Claude Code (recommended). Open Claude Code. Run these two commands one after the other:
// inside claude code
Done. The skill is installed globally for every Claude Code session. You'll notice Claude starts asking clarifying questions and explaining its plan before touching any code.
Path B · If you use any other tool (Cursor, Claude Desktop, Lovable, anything). Open a terminal, go to your project folder, and run this:
// inside your project folder
That command downloads the file and saves it as CLAUDE.md in your current folder. Claude (and Cursor) will automatically pick it up the next time you start a session in that project.
What You Just Installed.
1. Think Before Coding.Before Claude writes a single line, it has to state its assumptions out loud. "I'm assuming you want X. I'm assuming Y. Here's the tradeoff between A and B." No more silent decisions you find out about three hours later.
2. Simplicity First.Claude has to write the smallest amount of code that solves the actual problem. No defensive checks for situations that don't exist. No abstractions "in case" you need them later. The boring, obvious solution is the right one.
3. Surgical Changes.Claude can only edit what's necessary. Your existing style stays. Your existing code stays. If something is working, Claude doesn't touch it "while it's in there." This rule alone removes most of the headaches.
4. Goal-Driven Execution.Claude defines what "done" looks like before starting, then keeps going until it's actually done. No more "here's the code, good luck" when it's only 60% finished. Claude tests its own work against the goal.
The Difference Shows Up Immediately.
Easiest way to see the change. Pick any project (real one, demo one, doesn't matter). Open Claude. Paste this:
Watch what happens. Without the file, Claude usually starts writing code immediately. With the file, it lays out its assumptions, asks 1-2 clarifying questions ("do you want email-based reset or SMS?"), and waits for your green light. That's the magic — and you'll feel it on the very first task.
Add Rules That Match Your Stack.
The default file is intentionally generic. Karpathy's rules are universal. But your project has its own rules and you should add them. Open CLAUDE.md in any text editor and append your own at the bottom.
Beginner-friendly examples to add:
- "Always show me the plan in plain English before writing any code." Forces Claude to explain itself the way a friend would.
- "Never install new libraries without asking first." Stops surprise dependencies appearing in your project.
- "Use Tailwind for all styling. No inline CSS, no styled-components." Or swap in your own stack rules.
- "If you change more than one file, list every file you touched and why at the end." A receipt every time you ship.
Save the file. Restart Claude. Your rules now run alongside Karpathy's.
Bookmark These.
- The repo I posted about (Multica AI): github.com/multica-ai/andrej-karpathy-skills
- The original Claude Code plugin (Forrest Chang): github.com/forrestchang/andrej-karpathy-skills
- The raw CLAUDE.md file (direct download): raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md
- Karpathy's X post that started it all: x.com/karpathy
- Download Claude Code (if you don't have it): claude.com/download
Sixty-Five Lines. One File. Way Better Claude.
That's the whole upgrade. Drop in the file, watch Claude start asking better questions, writing less code, touching fewer files, and finishing what it starts. Stack your own rules on top and it gets even better.
Send me a screenshot of your first session with it. @tenfoldmarc.