forrestchang

andrej-karpathy-skills

AI#LLM #Claude #Prompt Engineering#Coding Agents

1005

// summary

This project provides a structured set of guidelines designed to improve LLM coding behavior by addressing common pitfalls like overcomplication and making wrong assumptions. It implements four core principles—Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution—to ensure more precise and verifiable code generation. Users can integrate these rules into their development workflow via a Claude Code plugin, a CLAUDE.md file, or Cursor project rules.

// technical analysis

This project provides a structured set of guidelines designed to mitigate common LLM coding pitfalls, such as over-complication, making unauthorized assumptions, and performing unnecessary refactoring. By implementing four core principles—Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution—the framework forces the AI to prioritize verifiable success criteria and minimal, intentional code changes. This approach shifts the LLM's behavior from impulsive execution to a more deliberate, goal-oriented process, effectively reducing technical debt and improving the quality of AI-generated code.

// key highlights

Promotes explicit reasoning by requiring the LLM to state assumptions and ask for clarification before implementation.

Enforces a 'Simplicity First' philosophy to prevent the common LLM tendency toward over-engineering and bloated abstractions.

Ensures 'Surgical Changes' by restricting the AI to modify only the code strictly necessary for the requested task.

Implements 'Goal-Driven Execution' by transforming vague instructions into test-based success criteria for better verification.

Provides flexible installation options via a Claude Code plugin or a standard CLAUDE.md file for project-specific integration.

Includes dedicated support for Cursor IDE through project-specific rule files to ensure consistent behavior across different development environments.

// use cases

Standardizing LLM coding behavior to prevent overengineering and unnecessary refactoring

Implementing goal-driven execution loops that prioritize test-based verification over imperative instructions

Enforcing surgical code changes to ensure that edits remain strictly scoped to user requests

// getting started

To begin, you can either install the guidelines as a Claude Code plugin using the marketplace command '/plugin install andrej-karpathy-skills@karpathy-skills' or download the CLAUDE.md file directly into your project root. If using Cursor, you can copy the provided .mdc rule file into your .cursor/rules directory to enforce these guidelines automatically. Once installed, the AI will reference these principles to guide its decision-making process during coding tasks.