Andrej Karpathy Skills Review: Worth Installing? (2026)

The most-starred "skill" in our directory is 358 words long. Andrej Karpathy Skills sits at roughly 217,000 GitHub stars, more than Anthropic's own official skills repo, and the core of it is one CLAUDE.md file you could read in under two minutes.

A famous name can collect stars for a page of common sense, so I checked. I've read every line, checked how the plugin is wired, and run an adapted version of these rules on the agent setup that maintains this site since April. Short version: install it, but know what you're getting. It's a behavior file, not a tool.

Quick verdict

Where it comes from

Karpathy didn't write this repo. In January 2026 he posted on X about the failure patterns he kept seeing when coding with LLMs: models make wrong assumptions and run with them without checking, and they overcomplicate code and bloat abstractions when a simpler fix exists. Developer forrestchang turned those complaints into instructions and published them on January 27, 2026. The repo has since moved to the multica-ai organization; old links still redirect.

That matters for how you read the name. "Andrej Karpathy Skills" is a community file inspired by his posts, not something he maintains or endorses. The SKILL.md links to the original post, which is the honest way to handle it.

What's inside

The repository has a handful of files:

• CLAUDE.md, the main guidelines
• skills/karpathy-guidelines/SKILL.md, the same content as an agent skill
• .claude-plugin/ with plugin.json and marketplace.json, so Claude Code can install it as a plugin
• .cursor/rules/karpathy-guidelines.mdc plus a CURSOR.md setup note
• EXAMPLES.md with before-and-after cases, and READMEs in English and Chinese

The guidelines open with an honest tradeoff line: they bias toward caution over speed, and for trivial tasks you should use judgment. Then come four principles.

1. Think Before Coding

State assumptions. If a request has two readings, present both instead of picking one silently. If there's a simpler approach, say so and push back. If something is unclear, stop and ask.

This is the rule that changes the most behavior. Without it, a model asked to "add caching" will choose a cache layer, an eviction policy and a TTL, then present the result as if those were your decisions.

2. Simplicity First

No features beyond the request, no abstractions for single-use code, no configurability nobody asked for, no error handling for impossible cases. The file's test: would a senior engineer call this overcomplicated? One line I like: if you wrote 200 lines and it could be 50, rewrite it.

3. Surgical Changes

Don't "improve" neighboring code, comments or formatting. Match existing style even if you'd write it differently. Mention unrelated dead code, but don't delete it. Clean up only the orphans your own change created. The closing test is strict: every changed line should trace back to the user's request.

4. Goal-Driven Execution

Turn vague tasks into checkable goals. "Fix the bug" becomes "write a test that reproduces it, then make it pass." Multi-step work gets a short plan where every step carries a verify: check. The point is to let the agent loop on its own without asking you whether it's done.

How to install it

You have three options, all verified against the repo README.

As a Claude Code plugin (applies everywhere):

/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills

The marketplace is named karpathy-skills in marketplace.json, which is why the install target ends in @karpathy-skills. If plugin marketplaces are new to you, our plugin marketplace explainer covers how they resolve.

As a project CLAUDE.md (new project):

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

Appended to an existing CLAUDE.md:

echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.md

The plugin route installs it as a skill, so Claude loads the full text only when the task looks like coding. That's the progressive disclosure model: the description sits in context, the body loads on demand. The CLAUDE.md route puts all 358 words in front of the model on every turn. For rules this short, I prefer the CLAUDE.md route. Behavior rules work best when they're always on, and the token cost is tiny.

If your team uses several agents, consider putting the text in an AGENTS.md instead. We compared the two formats in AGENTS.md vs CLAUDE.md.

What changed when we used it

The system that runs bestaifor.me's routines has carried a "Karpathy-adapted" coding section in its CLAUDE.md since April 30, 2026. We kept the four principles nearly word for word and added one rule of our own about verifying the final output before handing it over. Five months in, here's what I can say with confidence:

Diffs got smaller. The surgical-changes rule cut the habit of reformatting files the agent happened to open. Review got faster because the diff showed only the requested change.

More questions up front. The agent now asks "do you mean X or Y?" before starting, more often than before. That's the intended effect. It's also mildly annoying on simple tasks, which is exactly what the file's own tradeoff line warns about.

Over-building didn't vanish. Long sessions still drift. Once the context fills with earlier work, the agent sometimes adds a helper "for later." The rules reduce this; they don't eliminate it. A behavior file can't beat a model's habits every time.

I don't have a controlled before-and-after benchmark, and nobody else has published one that I trust. Treat the claims above as one team's experience, not a measurement.

Weaknesses

It's short on purpose, and that cuts both ways. There's no guidance on testing frameworks, commit messages, security or language conventions. You'll still need your own project rules alongside it.

The star count overstates it. 217,000 stars says more about Karpathy's name and timing than about how much the file does. Compare it with the Writing Skills meta-skill from Superpowers or the spec-driven workflow in Spec Kit, which do far more work per install. We put Spec Kit against two rivals in our spec-driven skills comparison.

Caution slows small tasks. "Ask when unclear" turns into extra round trips when you just want a typo fixed. You can soften this by adding one line: "For edits under 10 lines, proceed without asking."

No updates since April. That's fine for a text file, since there's nothing to break. It does mean the guidance won't adapt to newer agent features like subagents or hooks.

Security notes

This is about as safe as a skill gets. There are no executable scripts, no hooks, no MCP server, no network calls. The plugin manifest points at one skill folder. Still, read any file before it lands in your instructions. Our pre-install audit checklist takes five minutes, and for a plain-text skill like this one, reading the 358 words is the whole audit.

One caution on copycats: the popularity has spawned forks and lookalike repos. Install from the URL above, not from a search result.

How it compares to similar skills

These don't conflict. Karpathy's rules and Caveman stack well: one shapes what the agent does, the other shapes how much it says. Our roundup of the best agent skills for developers lists more, and our full list of AI tools for developers covers the editors these skills run inside.

Who should install it

Install it if your agent regularly hands you 300-line answers to 30-line problems, refactors code you didn't mention, or makes silent assumptions that you only discover in review. Those are the exact failures it targets.

Skip it if you already maintain a detailed CLAUDE.md with these ideas in it. You'll gain little, and duplicate rules waste context. Read it anyway and steal the "every changed line should trace to the request" test. That one sentence is the best thing in the file.

FAQ

Did Andrej Karpathy create this skill?

No. Developer forrestchang wrote it in January 2026, based on Karpathy's posts on X about LLM coding mistakes. The repo now lives under the multica-ai organization. Karpathy doesn't maintain it.

Is the Karpathy CLAUDE.md a skill or a CLAUDE.md file?

Both. The repo ships the same guidelines as a root CLAUDE.md, as a skill at skills/karpathy-guidelines/SKILL.md for the Claude Code plugin install, and as a Cursor rule file.

Does it work in Cursor or Codex?

Yes for Cursor, which has a dedicated .cursor/rules/karpathy-guidelines.mdc file. For Codex and other agents that read AGENTS.md, paste the CLAUDE.md text into your AGENTS.md. It's plain markdown, so any agent that reads instruction files can use it.

Will it slow Claude Code down?

Slightly, on small tasks. The rules tell the agent to ask before guessing, which adds a round trip when a request is ambiguous. On larger tasks it usually saves time because there's less to undo.

Is it safe to install?

Yes. It contains no scripts, hooks or network access, only markdown. Install from the official GitHub URL to avoid lookalike forks.