What is andrej-karpathy-skills? The four CLAUDE.md rules, and what installing them changes

"Karpathy skills" is the repository `multica-ai/andrej-karpathy-skills`: four plain-English rules for coding agents — Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution — packaged as a `CLAUDE.md`, a Claude Code plugin skill and a Cursor rule. It was not written by Andrej Karpathy. It was published by GitHub user forrestchang the day after Karpathy's January 2026 post about Claude coding (the repository was created on 27 January 2026 UTC), and it turns that post's complaints into instructions. Installing it adds text to your agent's context; it installs no scripts, hooks or commands.

We read the repository on 9 October 2026. GitHub showed 217,526 stars and 21,915 forks, last pushed on 20 April 2026. Stars show attention, not quality or safety.

Why people are searching for it

On 9 October 2026, about 09:00 Asia/Riyadh, Google Trends (worldwide, past 7 days) listed `karpathy skills` at +50% among rising queries related to claude skills. Trends values are relative, not search counts. Google's US English autocomplete the same morning suggested "andrej karpathy skills how to install", "…how to use", "…claude plugin" and "…vs ponytail".

On X, posts containing "karpathy skills", "andrej-karpathy-skills" or "karpathy-guidelines" (reposts excluded) numbered 8 over the last seven days. Search interest is rising; X chatter is light. The other rising names in the same Trends list — ponytail, impeccable, taste-skill, graphify — we covered between 5 October and 8 October, so this guide takes the one not yet on Skill Rank.

Where the four rules come from

Karpathy's post on X (26 January 2026), which we read today, lists the mistakes he saw: "The most common category is that the models make wrong assumptions on your behalf and just run along with them without checking." He adds that models "really like to overcomplicate code and APIs", "don't clean up dead code after themselves", and "still sometimes change/remove comments and code they don't like or don't sufficiently understand as side effects". He also wrote: "Don't tell it what to do, give it success criteria and watch it go."

One line in the same post is worth keeping in mind before you install: "All of this happens despite a few simple attempts to fix it via instructions in CLAUDE.md." The repository is a more careful version of that kind of instruction. It is a reasonable thing to try, not a proven fix.

What the four rules say

The file also states its own trade-off: "These guidelines bias toward caution over speed. For trivial tasks, use judgment." Expect more clarifying questions from the agent. That is the point, and it can feel slower on small edits.

What gets installed, by method

The repository has no install script. Each method copies or loads text.

Two practical notes:

• The commands still say `forrestchang`. The repository moved to multica-ai; on 9 October 2026 github.com/forrestchang/andrej-karpathy-skills redirected there and the raw CLAUDE.md URL still returned the file.
• The first `curl -o` replaces your file. If your project already has a CLAUDE.md, use the append form or merge by hand.

The repository's CURSOR.md says "Cursor does not read .claude-plugin/ or CLAUDE.md by default", which is why it ships a separate .mdc rule.

Before you install: three checks

1. License. The SKILL.md front matter, plugin.json and the README all say MIT, but the repository has no LICENSE file, and GitHub's API reported no detected license on 9 October 2026. For personal use this rarely matters. For a company repo, ask before copying.
2. Conflicts with your own rules. "Don't refactor things that aren't broken" will fight an instruction like "always modernise code you touch". Decide which wins and write it down.
3. `alwaysApply: true` in Cursor. The rule loads into every chat in that project. If you want it only for some tasks, change it to an agent-requested rule.

Karpathy guidelines vs ponytail

The autocomplete pairing is understandable, but the two do different jobs. Karpathy guidelines is a single always-on text that shapes how the agent behaves. Ponytail, as we described on 5 October, is a plugin with lifecycle hooks, so it runs code at points in the agent's session. Check that the two do not give conflicting instructions before using both.

FAQ

Did Andrej Karpathy write this? No. The plugin manifest lists the author as forrestchang, and the README says it was "derived from Andrej Karpathy's observations". The quotes come from his post.

Is it a skill or a CLAUDE.md? Both. The same four rules are shipped as CLAUDE.md, as skills/karpathy-guidelines/SKILL.md (used by the plugin), and as a Cursor .mdc rule.

Does it work with Codex or other agents? CURSOR.md suggests copying CLAUDE.md for tools that "only support a root instruction file". The content is plain Markdown, so any agent that reads an instructions file can use it.

Limits of this guide

• Based on the repository files, GitHub API data and Karpathy's X post, read on 9 October 2026. We did not install the plugin or measure any change in agent behaviour.
• Google Trends figures are relative. X counts depend on the query used.

Find it on Skill Rank as karpathy-guidelines, or browse coding skills.