The file with more than 200k stars is 2 KB and four rules in total. Karpathy did not write it, it was derived from a single post of his
You drop a CLAUDE.md into your project root and four of Claude Code's common mistakes get less frequent: making assumptions without asking, writing a simple task in a complicated way, touching code you never asked it to touch, and calling the job done without verifying it. The whole file is 2 KB, so it reads in two or three minutes. I verified three things and all three affect how you install it. First the name: the repo carries Karpathy's name but Karpathy did not write it, it was derived from complaints in one of his posts, and its own README says so. Second, the commands in the README point at the repo's old address; they work because GitHub redirects, but knowing the current one is safer. Third, and this is the one that hurts: the first command it suggests deletes your existing CLAUDE.md without asking. Below: the verified install, a way to set it up for every project at once instead of one, and the limit the file itself admits to.
The file itself
The repo does one thing: it gives you a CLAUDE.md that tells Claude Code how to behave. It holds four rules, all written around what the model should not do. It is MIT licensed and as of 19 September 2026 it has 214k stars and 21k forks. Its last update was 20 April 2026, so the file has been unchanged for five months. That is not a gap: there is no code in it that needs maintenance, just four paragraphs of behavioural rules.
What is extra below: what the four rules actually say, where the Karpathy in the name really stands, why the install command in the README can delete your existing CLAUDE.md, and how to set the file up for all your projects at once instead of just one.
What the four rules say
Each rule and the behaviour it blocks
| Rule | What it blocks |
|---|---|
| Think before coding | Assuming instead of asking. Silently picking one reading when several exist. Hiding its own confusion. |
| Simplicity first | Adding features nobody asked for, writing abstractions for single-use code, handling errors that cannot happen. If you wrote 200 lines and 50 would do, rewrite it. |
| Surgical changes | Improving adjacent code, comments or formatting. Refactoring what is not broken. If it spots dead code it should mention it, not delete it. |
| Goal-driven execution | Working without a success criterion. It asks the model to turn fix the bug into write a test that reproduces it, then make it pass, so it can verify itself. |
In the file's own words: these guidelines bias toward caution over speed. For trivial tasks, use judgment. So the full ritual of four rules is overkill when you are fixing a typo, and the file admits that itself.
About the Karpathy in the name
Not Karpathy's repo
The repo does not belong to Andrej Karpathy, it just carries his name. The first line of the README says so plainly: Karpathy-Inspired. The source is a single post where Karpathy listed how language models fail at coding; the person who wrote the file is someone else. That does not make the repo bad, its name just gets misread.
What Karpathy said in that post is the backbone of the file: models make wrong assumptions on your behalf and run along without checking, they do not manage their confusion, they build a thousand lines where a hundred would do, and they change code and comments they do not understand as a side effect. The four rules map one to one onto those four complaints.
Installing it: two routes
The first route is the plugin. It installs inside Claude Code and applies across all your projects, so you are not copying a file into each one. Two commands, both typed inside Claude Code, not in your terminal.
/plugin marketplace add multica-ai/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skillsThe karpathy-skills on the right of that second command is not a typo. The left side is the plugin name, the right side is the marketplace it comes from. They are two different things.
The second route is downloading the file straight into the project. That is enough if you only want to try it in one place. But the first command the README gives is dangerous, see the box below.
If you already have a CLAUDE.md
curl -o CLAUDE.md ... overwrites the file. If your project already has a CLAUDE.md, every project rule inside it is gone without a prompt.
One angle bracket overwrites, two append. If you already have the file, use the command with two.
# if you do NOT have a CLAUDE.md yet
curl -o CLAUDE.md https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md
# if you DO have one, this appends to it
echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.mdThe commands in the README say forrestchang, the ones above say multica-ai. Both point at the same repo: the project moved to an organisation account and the README was never updated. The old address still works thanks to GitHub's redirect, I tested both. Still, use the current address rather than leaning on a redirect.
Not one project, all of them
If you skipped the plugin route and do not want to copy the file into every project by hand, Claude Code has a personal instructions file. It sits in the .claude folder in your home directory and is read in every project. Append the file there and the rules apply everywhere.
# macOS and Linux
curl https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md >> ~/.claude/CLAUDE.md
# Windows PowerShell
curl.exe https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md >> $HOME\.claude\CLAUDE.mdTwo angle brackets here as well, so it appends. If you keep your own rules in that file, they stay.
Without overselling it
This file does not make the model smarter. What it does is narrower: it reduces four common mistakes. That is the repo's own measure too. You know it is working, it says, when unwanted changes in your diffs go down, when you rewrite less code that came back overcomplicated, and when its clarifying questions arrive before the implementation rather than after the mistake.
There is a Cursor side too: the repo ships a ready Cursor rule, so the same four rules apply there. The details are in CURSOR.md inside the repo.
DOA: Yapay Zeka ve Otomasyon
Installing these tools on your own is one thing; actually building with them is another. The community has people using these daily and people building systems from scratch.
If you want a system that actually runs in your business, let's talk for 10 minutes; I'll look at what you're trying to build and tell you which path fits. Free, and not a sales pitch.
Book a 10-minute callIf you'd rather learn this alongside people doing the same work instead of on your own, the community is always open:
Join the communityThis is an affiliate link.