karpathy-guidelines Skill for Claude Code: Simpler Code

Makes Claude state its assumptions, write the minimum code, touch only what it must and check its own work.

Installs: 43.9K; on skills.sh
Stars: 227.9K; on GitHub
Updated: Apr 2026; last repo commit

karpathy-guidelines gives Claude four rules for coding work, drawn from Andrej Karpathy's observations on common LLM coding mistakes. Think before coding: state assumptions and ask when unsure. Keep it simple: no features, abstractions or flexibility nobody asked for. Make surgical changes: every changed line traces to the request. Define success criteria and loop until verified.

It is one short SKILL.md with no scripts. It triggers when writing, reviewing or refactoring code. SKILL.md says it biases toward caution over speed, so for trivial tasks use judgment.

Good for

Developers tired of bloated diffs and drive-by refactors from the agent
Reviewers who want every changed line to trace to the request
Teams that want a short behaviour rulebook for Claude Code

Install karpathy-guidelines

Claude Code

Copy this prompt.

Install the karpathy-guidelines skill in Claude Code. Run this command:

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

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Pick one route. The README still writes the marketplace as forrestchang/andrej-karpathy-skills, the repo's old owner name. GitHub redirects that name to multica-ai/andrej-karpathy-skills, the repo we crawled and the one skills.sh lists, so both are the same project. The plugin manifest lists only this one skill and registers no hooks or MCP servers. The README also offers a per-project CLAUDE.md with the same rules; use that or the plugin, not both.

From the project README. Source.

What's in the skill

When it loads

Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.

Sections in SKILL.md

1. Think Before Coding
2. Simplicity First
3. Surgical Changes
4. Goal-Driven Execution

Key facts

Author: forrestchang, per the plugin manifest. Published in the multica-ai repo
Repo: multica-ai/andrej-karpathy-skills
Installs on skills.sh: 43,884 (Nov 2, 2026)
GitHub stars: 227,939 (Nov 2, 2026)
License: MIT
Last commit: Apr 20, 2026
Files: 1 file including SKILL.md
Source of the rules: Andrej Karpathy's observations on LLM coding pitfalls, linked from SKILL.md
skills.sh checks: Gen Agent Trust Hub pass, Socket pass, Snyk pass

What it includes

State assumptions, present interpretations, push back, stop and ask when unclear
SKILL.md: "Don't assume. Don't hide confusion. Surface tradeoffs."
1. Think Before Coding
Minimum code that solves the problem; no speculative features or abstractions
If 200 lines could be 50, rewrite it. The test is whether a senior engineer would call it overcomplicated.
2. Simplicity First
Touch only what you must; clean up only your own mess
Match existing style, mention unrelated dead code instead of deleting it, and remove only the orphans your change created.
3. Surgical Changes
Turn tasks into verifiable goals and loop until they pass
For example, "Fix the bug" becomes "Write a test that reproduces it, then make it pass." Multi-step work gets a short plan with a check per step.
4. Goal-Driven Execution

From the official docs

Questions

How do I install karpathy-guidelines?

Use the install box above. The README's recommended route is a Claude Code plugin, and it also offers a per-project CLAUDE.md file and a Cursor rule. The skills CLI line installs only the skill. The README writes the repo as forrestchang/andrej-karpathy-skills, an old name that GitHub redirects to multica-ai/andrej-karpathy-skills.

Does it change what Claude does, or is it just advice?

It is instructions that Claude reads when it writes, reviews or refactors code. It asks Claude to state assumptions, keep changes minimal and verify results. How closely Claude follows them depends on the task, and the README suggests judging by smaller diffs and fewer rewrites.

Will it slow Claude down?

It can on non-trivial work. SKILL.md says the guidelines bias toward caution over speed and that for trivial tasks you should use judgment. The README says the goal is fewer costly mistakes on non-trivial work, not slowing simple tasks.

Is this from Andrej Karpathy?

The rules are derived from Karpathy's observations, and SKILL.md links his post. The repo is published under multica-ai and the plugin manifest names forrestchang as author, not Karpathy.

Does karpathy-guidelines work in the Claude app or Claude.ai?

You can zip the skill folder and upload it under Customize, then Skills, in the Claude app. That is our generic instruction; the README only documents Claude Code, a CLAUDE.md file and a Cursor rule.

Does it need an API key, and is it safe?

Neither SKILL.md nor the README names an API key. The skill bundles no scripts; it adds instructions that Claude follows while it writes code, and the edits Claude then makes go through your normal permission prompts.

Skill or CLAUDE.md?

The README offers both. The plugin installs the skill for all your projects. Option B copies the same guidelines into a project CLAUDE.md. Pick one so the rules are not repeated.

Information

updated Nov 2, 2026

Repository: multica-ai/andrej-karpathy-skills
Folder: skills/karpathy-guidelines

multica-ai/andrej-karpathy-skills skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 5

andrej-karpathy-skills · github.com
README.md · github.com
plugin.json · github.com
commands.md · github.com
SKILL.md · github.com