Vibeleaderboard
← Back to Tools
Index / tool

Andrej Karpathy Skills

github.com/forrestchang/andrej-karpathy-skills
Visit github.com
Category
AI Tools
Rank
No. 74Tools index
Pricing
Open Source
Type
TOOL
Builder
forrestchang
GitHub
211.0k stars
Date

About

A structured set of coding guidelines for Claude AI that addresses common LLM pitfalls like overcomplication, unnecessary changes, and poor assumption management. Based on Andrej Karpathy's observations about how to get better code output from AI assistants.

What it does

It gives coding agents four operating rules: clarify uncertainty, prefer the smallest adequate solution, limit edits to the requested scope, and define checks before implementation. The guidance can apply globally through Claude Code, per project through root instructions, or automatically as a Cursor rule.

Why it's ranked here

This is a focused, practical behavior layer with unusually concrete boundaries. It tells agents when to ask, when to push back, what not to modify, and how to turn vague work into testable outcomes. Its value comes from disciplined scope control, not added technical capability.

What's good

The rules translate recurring coding failures into observable checks. Every changed line should trace to the request. Bug fixes should begin with reproduction tests. Refactors should preserve passing tests. The included examples contrast bloated or invasive responses with smaller alternatives, making the intended behavior easy to teach and audit.

Tradeoffs

The guidance deliberately favors caution over speed, so trivial edits may attract needless questions or planning. Its simplicity rules can also discourage abstractions or defensive handling that real requirements justify. It provides behavioral instructions, not enforcement, code execution, domain knowledge, or proof that an agent followed them.

How to use it well

Use it with coding agents handling bug fixes, refactors, and ambiguous feature requests where diff discipline matters. Add project-specific conventions, architecture constraints, and required test commands alongside it. Apply judgment for obvious one-line changes. It does not replace technical specifications, repository documentation, automated tests, or human review.

Technical notes+

CLAUDE.md contains the four core principles for per-project use. skills/karpathy-guidelines/SKILL.md packages substantially the same text as a reusable skill, while .claude-plugin/plugin.json and .claude-plugin/marketplace.json expose it through the Claude Code plugin marketplace. .cursor/rules/karpathy-guidelines.mdc sets alwaysApply: true; CURSOR.md explains copying that rule or installing the personal skill. EXAMPLES.md supplies worked Python and diff examples. README.md documents plugin and direct-file installation, and README.zh.md provides Simplified Chinese documentation. The contributor guidance explicitly requires synchronized edits across the core instruction, Cursor rule, and published skill text.

Observed

License
MIT
Content format
Markdown guidance with JSON plugin metadata
Install surfaces
Claude Code marketplace plugin, per-project root instructions, Cursor project rule, or personal Cursor skill
Interfaces
Agent instructions and editor rules; no library, API, MCP server, or runtime CLI is described
Editor support
Claude Code and Cursor
Localization
English and Simplified Chinese documentation
Structure
One reusable skill, one always-applied Cursor rule, plugin manifests, and a worked examples document

Read from README.md, CLAUDE.md, CURSOR.md, EXAMPLES.md, README.zh.md, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, skills/karpathy-guidelines/SKILL.md, .cursor/rules/karpathy-guidelines.mdc.

What it can do

  • Implement Think Before Coding principle

    Coding request or problem statement → Structured analysis before code generation

  • Apply Simplicity First guidelines

    Complex coding requirements → Simplified code solutions avoiding overcomplication

  • Execute Surgical Changes methodology

    Existing codebase requiring modifications → Minimal, targeted code changes

  • Enforce Goal-Driven Execution

    Vague or broad coding objectives → Focused, goal-aligned code implementation

  • Manage coding assumptions

    Ambiguous development requirements → Clarified assumptions and constraints

  • Configure Claude AI coding behavior

    CLAUDE.md file or plugin installation → Enhanced AI assistant with improved coding practices

Tags

claudeai-guidelinescodingllmpromptsdeveloper-toolskarpathy

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.