Karpathy Guidelines

Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM coding pitfalls.

Tradeoff: These guidelines bias toward caution over speed. For trivial tasks, use judgment.

1. Think Before Coding

Don't assume. Don't hide confusion. Surface tradeoffs.

Before implementing:

• State your assumptions explicitly. If uncertain, ask.
• If multiple interpretations exist, present them - don't pick silently.
• If a simpler approach exists, say so. Push back when warranted.
• If something is unclear, stop. Name what's confusing. Ask.

2. Simplicity First

Minimum code that solves the problem. Nothing speculative.

• No features beyond what was asked.
• No abstractions for single-use code.
• No "flexibility" or "configurability" that wasn't requested.
• No error handling for impossible scenarios.
• If you write 200 lines and it could be 50, rewrite it.

Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

3. Surgical Changes

Touch only what you must. Clean up only your own mess.

When editing existing code:

• Don't "improve" adjacent code, comments, or formatting.
• Don't refactor things that aren't broken.
• Match existing style, even if you'd do it differently.
• If you notice unrelated dead code, mention it - don't delete it.

When your changes create orphans:

• Remove imports/variables/functions that YOUR changes made unused.
• Don't remove pre-existing dead code unless asked.

The test: Every changed line should trace directly to the user's request.

4. Goal-Driven Execution

Define success criteria. Loop until verified.

Transform tasks into verifiable goals:

• "Add validation" → "Write tests for invalid inputs, then make them pass"
• "Fix the bug" → "Write a test that reproduces it, then make it pass"
• "Refactor X" → "Ensure tests pass before and after"

For multi-step tasks, state a brief plan:

1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]

Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.

Karpathy-Inspired Agent Skills

Karpathy-inspired coding-agent guidelines packaged for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Aider, GitHub Copilot, OpenClaw, Warp, Windsurf, Cline, and the wider SwarmVault/SwarmClaw agent matrix.

The canonical source is skills/karpathy-guidelines/SKILL.md. Agent-specific files are generated under adapters/ so the repository root stays readable while every supported tool receives the same behavioral guidance.

English | Simplified Chinese

What This Adds

These guidelines address four common LLM coding failure modes described by Andrej Karpathy:

Principle	Prevents
Think Before Coding	Silent assumptions, hidden confusion, missing tradeoffs
Simplicity First	Over-engineering, bloated abstractions, speculative features
Surgical Changes	Drive-by rewrites, unrelated cleanup, accidental behavior changes
Goal-Driven Execution	Vague completion criteria and unverified changes

The content is intentionally short. It is meant to merge with project-specific instructions, not replace them.

Quick Install

Install through npm and run the installer with npx:

npx @swarmclawai/andrej-karpathy-skills --list
npx @swarmclawai/andrej-karpathy-skills --agent codex --dest /path/to/project

Or install the CLI globally:

npm install -g @swarmclawai/andrej-karpathy-skills
andrej-karpathy-skills --agent cursor --dest /path/to/project

The CLI copies the right adapter from the npm package into the destination project.

Clone Install

Clone once, then copy the right adapter into any project. The installer reads install/targets.json, copies from adapters/<agent-id>/..., and writes to the real path your agent expects.

git clone https://github.com/swarmclawai/andrej-karpathy-skills.git
cd andrej-karpathy-skills
node scripts/install.mjs --list
node scripts/install.mjs --agent codex --dest /path/to/project

Use --force if you intentionally want to overwrite an existing instruction file:

node scripts/install.mjs --agent claude --dest /path/to/project --force
node scripts/install.mjs --agent openclaw --dest /path/to/project --force

Use --global for agents with supported user-level skill locations:

node scripts/install.mjs --agent claude --global --force
node scripts/install.mjs --agent hermes --global --force

Adapter Layout

If you only need one file, copy the adapter source to the install target shown below:

| Agent | Adapter source | Installs to | | --- | --- | | Codex | adapters/codex/AGENTS.md | AGENTS.md | | Claude Code | adapters/claude/CLAUDE.md | CLAUDE.md | | Cursor | adapters/cursor/.cursor/rules/karpathy-guidelines.mdc | .cursor/rules/karpathy-guidelines.mdc | | Gemini CLI | adapters/gemini/GEMINI.md | GEMINI.md | | Aider | adapters/aider/CONVENTIONS.md | CONVENTIONS.md | | GitHub Copilot | adapters/copilot/.github/copilot-instructions.md | .github/copilot-instructions.md | | VS Code Copilot Chat | adapters/vscode/.github/chatmodes/karpathy-guidelines.chatmode.md | .github/chatmodes/karpathy-guidelines.chatmode.md | | OpenClaw | adapters/openclaw/.openclaw/skills/karpathy-guidelines/SKILL.md | .openclaw/skills/karpathy-guidelines/SKILL.md | | Warp or shared Agent Skills | adapters/warp/.agents/skills/karpathy-guidelines/SKILL.md | .agents/skills/karpathy-guidelines/SKILL.md | | Windsurf | adapters/windsurf/.windsurfrules | .windsurfrules |

The full target matrix lives in install/targets.json.

Supported Agent Matrix

The matrix is based on SwarmVault's installer targets and SwarmClaw's agent provider registry, then cross-checked with public agent documentation where available.

Core targets:

codex, claude, cursor, goose, pi, gemini, opencode, aider, copilot, trae, claw, droid, kiro, hermes, antigravity, vscode.

Extended Agent Skills targets:

amp, augment, adal, bob, cline, codebuddy, command-code, continue, cortex, crush, deepagents, firebender, iflow, junie, kilo-code, kimi, kode, mcpjam, mistral-vibe, mux, neovate, openclaw, openhands, pochi, qoder, qwen-code, replit, roo-code, trae-cn, warp, windsurf, zencoder.

See docs/agent-compatibility.md for the research notes and source links.

Claude Code Plugin

This repo still includes Claude plugin metadata at the root because plugin tooling expects it there:

.claude-plugin/plugin.json
.claude-plugin/marketplace.json
skills/karpathy-guidelines/SKILL.md

When published as a Claude plugin marketplace entry, the plugin exposes the same canonical skill used by every other adapter.

Maintain The Package

Edit the canonical skill, then regenerate and verify:

npm run generate
npm run verify
npm run pack:dry-run

Generated adapters live under adapters/<agent-id>/ and start with a comment that names their source. Do not edit generated adapters by hand; update skills/karpathy-guidelines/SKILL.md or scripts/generate.mjs.

Attribution

This package is based on the original forrestchang/andrej-karpathy-skills repository and preserves the MIT license attribution.

License

MIT