Andrej Karpathy Skills¶

What it is¶

A curated collection of skills and patterns inspired by Andrej Karpathy's approach to AI and software engineering, designed to help agents avoid basic pitfalls.

What problem it solves¶

It codifies high-signal development habits and "instincts" into actionable patterns for AI agents.

Where it fits in the stack¶

Category: AI & Knowledge / Best Practices

Typical use cases¶

Agent Initialization: Setting baseline "thinking" patterns for a new project.
Workflow Optimization: Reducing agent hallucinations and "looping" by enforcing clarifying questions.
Code Review Standard: Using the guidelines as a checklist for human or AI code reviews.

Strengths¶

Low Overhead: Simple Markdown-based guidelines that don't require complex infrastructure.
High Signal: Focuses on the most common and damaging mistakes made by AI agents.
Developer-Centric: Aligns AI behavior with senior-level software engineering best practices.

Limitations¶

Opinionated: Some patterns might conflict with specific project styles or requirements.
Manual Enforcement: Outside of Claude Code, requires manual inclusion in system prompts or project context.

When to use it¶

When you find your AI agent is making "obvious" mistakes or over-complicating solutions.
At the start of a new project to establish a high bar for code quality.

When not to use it¶

For highly experimental or "creative" coding where standard constraints might be too restrictive.
In environments where a different, strict coding standard is already enforced.

Licensing and cost¶

Open Source: Yes (MIT)
Cost: Free
Self-hostable: N/A (Documentation/Guidelines)

Getting started¶

Installation¶

Karpathy-inspired guidelines can be installed as a plugin for Claude Code:

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

Alternatively, you can manually add the CLAUDE.md to your project:

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

Sample CLAUDE.md Snippet¶

## Karpathy Instincts
- **Think Before Coding**: Surface tradeoffs. Ask when unclear.
- **Simplicity First**: Minimal code. No speculative features.
- **Surgical Changes**: Touch only what you must. Match existing style.
- **Goal-Driven Execution**: Define success criteria. Loop until verified.

Basic usage¶

Once installed, the guidelines are automatically applied by the AI agent during its thinking process. You can verify it's working by observing if the agent: - Asks clarifying questions before implementation. - Prefers simple, direct solutions over complex abstractions. - Makes surgical changes to the code.

Thinking Pattern Example: Handling Ambiguity¶

When faced with a vague request like "Fix the auth," an agent using Karpathy/Pocock skills avoids jumping straight into code.

Agent Thinking Process: 1. Initial Impulse: Look for auth.py and start changing login logic. 2. Skill Check (Simplicity/Surgical): Wait. "Fix the auth" is ambiguous. I need more context to be surgical and avoid over-engineering. 3. Action: Ask the user: "Could you specify if this is a bug with the current login flow, an integration issue with the new provider, or a request for a new feature like MFA?" 4. Result: User specifies it's a redirect loop after OAuth. Agent now knows exactly which file and function to target, minimizing code churn.

Thinking Pattern Example: The "Zero-Draft" Pattern¶

Instead of writing 200 lines of complex boilerplate, the agent writes a "Zero-Draft" that proves the core logic works with minimum dependencies.

Step 1: Write a 10-line script that triggers the core API or logic.
Step 2: Verify the draft runs and returns the expected result.
Step 3: "Refactor" the proven logic into the production codebase surgically.

CLI examples¶

The skills can be managed via the Claude Code CLI:

# List installed plugins
/plugin list

# Update the skill
/plugin update andrej-karpathy-skills

Sources / references¶

Contribution Metadata¶

Last reviewed: 2026-05-11
Confidence: high