Turning Andrej Karpathy’s Coding Insights into a Claude.md Guide

Andrej Karpathy previously shared typical failure patterns of large language models when programming. A GitHub repository has transformed each of those observations into concrete behavior rules stored in a CLAUDE.md file.

The file lists four core principles: “think before coding” (addresses wrong assumptions), “simplicity first” (prevents over‑design), “surgical edits” (modify only necessary parts and clean up only the mess it caused), and “goal‑driven execution” (e.g., turn “add validation” into “write tests for invalid inputs and make them pass”).

For example, the surgical‑edit rule explicitly tells Claude not to “improve” adjacent code, comments, or formatting, nor to refactor code that is not broken.

Goal‑driven execution converts a vague instruction into a concrete test‑first loop, allowing the AI to iterate until the specified criteria are satisfied.

To use the guide, place CLAUDE.md in the root of any project; Claude automatically reads it at the start of each new session.

User feedback reported that the configuration makes a noticeable difference, with comments such as “it’s like night and day after using it.”

Repository: https://github.com/forrestchang/andrej-karpathy-skills