karpathy-guidelines

karpathy-skills-anycoding

Universal, reusable coding-agent skills for OpenCode, Claude Code, Cursor, Trae, OpenClaw, and other AI coding tools.

English | 简体中文

Install behavior, not boilerplate

Most AI coding tools can write code.

Far fewer can consistently:

• stop and ask when requirements are ambiguous
• choose the smallest correct implementation
• avoid unrelated edits
• define success criteria before changing code

This project gives coding agents a compact behavior layer that improves those habits across tools.

What this is

karpathy-skills-anycoding is a tool-agnostic skill pack that helps AI coding assistants behave more like pragmatic senior engineers.

It turns a small set of high-signal engineering behaviors into reusable instruction files, adapters, and install scripts that work across multiple coding-agent tools.

Why this is different from a random prompt

This repository is designed to be:

• portable across multiple AI coding tools
• short enough to stay influential
• practical enough for real repositories
• explicit about compatibility instead of overclaiming support

It is not a giant framework prompt. It is a small operating layer for coding behavior.

Why people use it

Many coding agents still tend to:

• make silent assumptions
• overengineer solutions
• edit unrelated code
• start coding before defining success criteria

This repository pushes them toward the opposite behavior:

• clarify ambiguity first
• prefer the smallest correct solution
• make surgical diffs
• verify outcomes against concrete checks

The four principles

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

These principles are inspired by Andrej Karpathy's public comments about common LLM coding failure modes.

Install in one command

Use the installer that matches your tool and run it inside the repository where you want the rules to apply.

If your tool supports project instructions, rules files, custom skills, or system prompts, this project gives you a ready-to-install starting point.

By default, the installer writes into your current working directory. For OpenCode, that means you should usually run it from your project root so it creates ./AGENTS.md for that project.

OpenCode

curl -fsSL https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.sh | bash -s -- --tool opencode

Claude Code

curl -fsSL https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.sh | bash -s -- --tool claude

Cursor

curl -fsSL https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.sh | bash -s -- --tool cursor

Universal install

curl -fsSL https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.sh | bash -s -- --tool universal --output AGENTS.md

PowerShell

irm https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.ps1 | iex
Install-KarpathySkills -Tool claude

What the installer does

• opencode appends a marked block into AGENTS.md
• claude appends a marked block into CLAUDE.md
• cursor writes .cursor/rules/karpathy-guidelines.mdc
• universal appends a portable block into a target file such as AGENTS.md

If the block already exists, the installer skips duplicate insertion.

OpenCode project vs global install

For OpenCode, there are two useful locations:

• Project-specific rules: ./AGENTS.md in the project root
• Global personal rules: ~/.config/opencode/AGENTS.md

Recommended default: install into the project root.

If you want global personal behavior across all OpenCode sessions, you can either copy the generated file manually:

cp AGENTS.md ~/.config/opencode/AGENTS.md

or install directly to the global path:

curl -fsSL https://raw.githubusercontent.com/Vincent-A-Yang/karpathy-skills-anycoding/anycoding/scripts/install.sh | bash -s -- --tool opencode --output ~/.config/opencode/AGENTS.md

Supported tools

Confirmed direct instruction surfaces

These have known file or rule entry points and should work directly when installed to the expected path:

• OpenCode via project-root AGENTS.md
• Claude Code via project-root CLAUDE.md
• Cursor via .cursor/rules/karpathy-guidelines.mdc

Portable adapter templates

These adapters are included as ready-to-use templates, but exact auto-loading behavior depends on the tool version and configuration model:

• Trae
• OpenClaw
• other prompt-driven or rules-driven coding agents

Common use cases

• make an AI coding assistant stop overengineering routine changes
• reduce drive-by refactors in generated diffs
• add a reusable behavior layer across multiple tools in one team
• improve project instruction quality without writing a long custom system prompt
• bootstrap better coding habits for a new agent-enabled repository

Quick start

Option 1: use the core skill

Copy core/karpathy-anycoding.md into your tool's instruction surface.

Option 2: use a ready-made adapter

Pick the closest file from adapters/ and place it where your tool expects project instructions.

Option 3: use the reusable skill package

If your tool supports a SKILL.md-style reusable package, use skills/karpathy-guidelines/SKILL.md.

What success looks like

You know the skill is working when you start seeing:

• more clarifying questions before implementation
• smaller diffs with fewer unrelated edits
• simpler code on the first pass
• more explicit verification and fewer vague completions

Repository structure

karpathy-skills-anycoding/
├── README.md
├── README.zh.md
├── LICENSE
├── CHANGELOG.md
├── CODE_OF_CONDUCT.md
├── CONTRIBUTING.md
├── .gitignore
├── core/
│   └── karpathy-anycoding.md
├── scripts/
│   ├── install.sh
│   └── install.ps1
├── skills/
│   └── karpathy-guidelines/
│       └── SKILL.md
├── adapters/
│   ├── claude/
│   │   └── CLAUDE.md
│   ├── cursor/
│   │   └── .cursor/
│   │       └── rules/
│   │           └── karpathy-guidelines.mdc
│   ├── opencode/
│   │   └── AGENTS.md
│   ├── openclaw/
│   │   └── AGENTS.md
│   ├── trae/
│   │   └── AGENTS.md
│   └── universal/
│       └── SYSTEM_PROMPT.md
└── docs/
    ├── ADOPTION.md
    ├── COMPATIBILITY.md
    └── GITHUB_METADATA.md

Recommended setup by tool

OpenCode

Use the installer or copy adapters/opencode/AGENTS.md to your project root as AGENTS.md.

For global personal defaults in OpenCode, place the same content at ~/.config/opencode/AGENTS.md.

Claude Code

Use the installer or copy adapters/claude/CLAUDE.md to your project root as CLAUDE.md.

Cursor

Use the installer or place adapters/cursor/.cursor/rules/karpathy-guidelines.mdc in .cursor/rules/.

Trae

Start from adapters/trae/AGENTS.md, then place it into Trae's workspace instruction or agent profile surface.

OpenClaw

Start from adapters/openclaw/AGENTS.md, then place it into the tool's project instruction or bootstrap prompt surface.

Other AI coding assistants

Start from adapters/universal/SYSTEM_PROMPT.md or core/karpathy-anycoding.md.

Why this approach works

This project focuses on behavioral instructions, not tool-private plugin logic.

That makes it portable across tools that support any of the following:

• project instructions
• rules files
• custom skills
• system prompts
• workspace memory
• agent profiles

It works best when combined with repository-specific engineering rules such as language conventions, testing requirements, framework patterns, and security constraints.

FAQ

Is this tied to one AI tool?

No. The core is tool-agnostic. Some adapters target known file-based entry points, while others are portable templates.

Does this