Pickandrej-karpathy-skills: 4 Principles to Reduce LLM Coding Mistakes
multica-ai/andrej-karpathy-skills
About the project
This is a guidelines file designed to address issues where LLMs hide incorrect assumptions or introduce unnecessary abstractions while writing code. It is designed to counter the LLM coding pitfalls pointed out by Andrej Karpathy. Provided as a single CLAUDE.md file, it can be applied directly to the Claude Code environment.
It contains four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. It encourages asking questions instead of guessing when requirements are ambiguous, and instructs the removal of unrequested features or error handling. It also restricts touching parts of the existing code unrelated to the task to reduce unnecessary diffs.
The core idea is to encourage setting verifiable success criteria instead of imperative instructions. For example, transforming 'fix bug' into 'write a reproduction test and pass it' allows the LLM to loop independently. A .cursor/rules file is also provided to apply the same rules in Cursor.
Excessive strictness is not required for trivial tasks such as simple typo fixes, and it is recommended to merge with project-specific rules. It is released under the MIT license and focuses on reducing costly mistakes in complex tasks.
multica-ai/andrej-karpathy-skills
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.