Karpathy CLAUDE.md: how to install it, and what to check first
The Karpathy-inspired CLAUDE.md repository: the plugin and the file route, the same text shipped three ways, CLAUDE.md or skill, and its missing license file.
- For
- Claude Code users who found the Karpathy-inspired CLAUDE.md
- Written for
- Claude Code
- Time
- 7 min read · updated Nov 21, 2026
The short answer
The “Karpathy CLAUDE.md” is a public repository whose README says it is one CLAUDE.md file derived from Andrej Karpathy’s observations about LLM coding mistakes, in four principles: Think Before Coding, Simplicity First, Surgical Changes and Goal-Driven Execution. It ships the same text three ways: as a plugin (/plugin marketplace add forrestchang/andrej-karpathy-skills, then /plugin install andrej-karpathy-skills@karpathy-skills), as a file you download into a project, and as a Cursor rule. Pick one. One catch to check before you copy it: the repository has no LICENSE file and GitHub detects none, although its plugin manifest says MIT.
Most people who search “karpathy claude md” have seen a post or a GitHub link about a CLAUDE.md file that makes Claude Code more careful, and want to know how to get it and whether to. This page reads the repository, its manifests and Claude Code’s own documentation on 21 November 2026. We did not install the plugin for this page, and we have no measurement of how the guidelines change Claude’s behavior, so it makes no claim about whether they work. It covers what is in the repository, the two ways to install it, how a CLAUDE.md differs from a skill, and the license question.
What the repository is
It is a small repository. GitHub’s API answers at forrestchang/andrej-karpathy-skills and returns multica-ai/andrej-karpathy-skills, so it has moved; the install commands in the README still use the old owner name. Its description says it is a single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy’s observations on LLM coding pitfalls. The README links to the post on X that it draws on; we could not read that post, so we do not describe it. The file itself is 65 lines. Its four sections are headed Think Before Coding, Simplicity First, Surgical Changes and Goal-Driven Execution. The last commit in our check was on 20 April 2026.
This page names the four headings and stops there. It does not quote or summarize what is under them, because of the license question below. Read the file at the repository to see it.
One text, three forms
The repository packages the same guidelines three ways. We compared the plugin’s skill file with CLAUDE.md: apart from the title lines, one introductory sentence, the skill’s frontmatter and a closing note that only the CLAUDE.md has, the text is the same. The third form is a Cursor rule, which does not apply to Claude Code.
Two ways to install it in Claude Code
The README calls the plugin route recommended, and offers the file as the per-project alternative. They are alternatives: the same guidelines twice is not more guidelines.
In a session, /plugin install opens the plugin’s details panel first, where you choose a scope; Best Claude Code plugins explains the panel. The README says the plugin makes the skill available across all your projects. Once installed it appears in / as andrej-karpathy-skills:karpathy-guidelines, and claude plugin list shows it. To remove it, claude plugin uninstall andrej-karpathy-skills@karpathy-skills.
For the file route the README gives two commands, one for a new project and one that appends to an existing file. Mind the difference: the first one replaces any CLAUDE.md already in that folder.
Both URLs point at the main branch, so what you download is whatever is there that day. Whatever lands in CLAUDE.md goes into your agent’s context at the start of every session, so read it before you keep it, and commit your copy so it stays the version you read. Claude Code’s docs say that running /context all in a session lists the memory files that loaded, which is how to confirm it worked.
CLAUDE.md or skill
Both are text Claude reads, and they differ in when. Claude Code’s memory docs say CLAUDE.md is read at the start of every session, and recommend keeping it to facts Claude should hold in every session: build commands, conventions, project layout, “always do X” rules. A procedure that matters only sometimes belongs in a skill or a path-scoped rule. A skill loads when it is relevant or when you call it. Skills vs prompts, commands and plugins sets the forms side by side.
By that description, behavior guidelines meant to apply to all your coding work fit the always-loaded form, and the plugin gives the on-demand form of the same text. Two things from the docs are worth knowing either way. Anthropic says to target under 200 lines per CLAUDE.md, because longer files use more context and reduce adherence, and this one is 65. And it says Claude treats CLAUDE.md as context, not enforced configuration: if something must always happen, such as running a check before every commit, the docs say to use a hook.
The docs also ask for instructions concrete enough to verify, with the examples “Use 2-space indentation” instead of “Format code properly” and “Run npm test before committing” instead of “Test your changes”. A file of principles says how to behave; your project’s own lines say what is true about your repository. They do different jobs, and you can have both.
That sketch is ours, written in the style the docs recommend; replace every line with your own. /init in Claude Code writes a first CLAUDE.md from your codebase.
The license question
The repository’s plugin manifest and its skill’s frontmatter both say MIT. The repository root has no LICENSE file, GitHub’s API reports no license for it, and the license endpoint returns not found. A declared MIT line in a manifest and no license file are two different signals, and we cannot tell you which one the author intends. If you are putting the file into a company repository, a client project or anything you redistribute, ask the author or write your own guidelines. This is a plain reading, not legal advice.
Skills that make a finished file
A CLAUDE.md tells the agent how to behave; it does not make anything you can open. Lookbk’s skills do that: each one renders a file, such as a landing page, a slide deck, a link-preview image or a title video, and its page shows what it made. They are plain skill folders and work next to any CLAUDE.md. These four are free.
Run the installer from inside Claude Code with the last part of a skill page’s address; it writes to ~/.claude/skills, the folder Claude Code reads.
Common questions
- What is the Karpathy CLAUDE.md?
- A public repository whose README says it is a single CLAUDE.md file for Claude Code, derived from Andrej Karpathy’s observations on LLM coding pitfalls, in four principles: Think Before Coding, Simplicity First, Surgical Changes and Goal-Driven Execution.
- How do I install the Karpathy CLAUDE.md in Claude Code?
- Either as a plugin:
/plugin marketplace add forrestchang/andrej-karpathy-skills, then/plugin install andrej-karpathy-skills@karpathy-skills. Or by downloading its CLAUDE.md into your project with the curl commands in the README. Use one route, not both. - Should it be a CLAUDE.md or a skill?
- Claude Code’s docs say CLAUDE.md is read at the start of every session and a skill loads when relevant. The repository ships the same text both ways, so it comes down to whether you want it always loaded or on demand.
- Is the Karpathy CLAUDE.md licensed for use?
- Its plugin manifest and skill frontmatter say MIT, but the repository has no LICENSE file and GitHub detects no license. Check with the author before using it in a company or client repository.
- Does it work in Cursor?
- The repository has a separate Cursor rule,
.cursor/rules/karpathy-guidelines.mdc, and its CURSOR.md says Cursor does not readCLAUDE.mdby default. Cursor rules vs skills explains how Cursor rules work. - How long should a CLAUDE.md be?
- Claude Code’s docs say to target under 200 lines per file, because longer files use more context and reduce adherence. The Karpathy file is 65 lines.
- How do I remove it?
- For the plugin,
claude plugin uninstall andrej-karpathy-skills@karpathy-skills. For the file, delete the lines you added from your CLAUDE.md.
Sources
What this guide says about other companies’ tools comes from their own documentation, checked on Nov 21, 2026.
Did this guide help?
Read next
All guidesClaude Code skills vs prompts, commands and plugins
What a skill is next to a prompt, a slash command, a subagent, a plugin, CLAUDE.md and a hook, with each definition taken from Anthropic’s Claude Code docs.
Claude Code skills best practices
How to choose, install, check, update and combine Claude Code skills, from Anthropic’s docs: where they live, how they trigger and what to do when they don’t.
Write a description that makes a skill trigger
How an agent picks a skill, what to put in its description, and how to test that it triggers when it should and stays quiet when it shouldn’t.
Safety checks before running a skill
A skill is instructions and code your agent may run. Six checks to make first: its source, its commands, its network use and what it may do without asking.
Claude Code plugins: install them, and the best ones
Claude Code plugins explained: browse the marketplace, install one with /plugin, pick a scope, then the best plugins by job, each checked at its source.
See a skill’s result before you install it.
Every skill page shows what it makes, what it needs and which agents can run it.



