karpathy-coding-principles

When NOT to Use

• Task is outside your authorization scope
• You need to implement controls (use implementing-* skills)
• Task is about analysis, not action (use analyzing-* skills)
• You don't have access to target systems
• Task requires compliance expertise (consult professionals)
• Task is about defense, not offense (use defensive skills)

Overview

Andrej Karpathy's 4 coding principles distilled into actionable rules for AI agents and developers. Focuses on reducing overengineering, maintaining simplicity, making minimal changes, and verifying outcomes. Originally from multica-ai/andrej-karpathy-skills.

Capabilities

• Enforce explicit reasoning before writing code
• Prevent speculative features and unnecessary abstractions
• Restrict changes to only what the task requires
• Convert imperative tasks into declarative goals with verification

When to Use

Trigger phrases:

• "karpathy coding principles"

• "Writing or reviewing code for quality"

• "Reducing overengineering in implementations"

• "Coaching LLMs to produce cleaner output"

• Writing or reviewing code for quality

• Reducing overengineering in implementations

• Coaching LLMs to produce cleaner output

• Code review to catch scope creep

• Refactoring sessions to simplify architecture

• Any coding task where discipline matters

Principle 1: Think Before Coding

Before writing any code, reason explicitly about the problem.

• Surface assumptions — state what you believe about the requirements
• Ask clarifying questions — do not guess when ambiguity exists
• Outline your approach in plain language before implementation
• Consider edge cases and failure modes upfront

# BAD: Jump straight to code
def process(data):
    # immediately writing implementation

# GOOD: Think first
# Assumption: data is a list of dicts with 'id' and 'value' keys
# Edge cases: empty list, missing keys, negative values
# Approach: validate -> transform -> aggregate
def process(data):
    ...