karpathy-guidelines

To prevent LLM coding mistakes during refactoring, apply behavioral guidelines that surface assumptions, enforce simplicity, and guide surgical changes. This ensures modifications remain minimal and existing behavior is preserved.

The best way to keep LLM generated code simple is to enforce a simplicity-first guideline that chooses minimal solutions and avoids unnecessary abstractions during the code generation and review process.

To make surgical code changes with an LLM without breaking existing behavior, apply guidelines that modify only necessary code while explicitly preserving all other existing functionality elsewhere in the codebase.

Your LLM coding task lacks verifiable success criteria because it requires goal-driven execution guidelines to define measurable outcomes and loop until the code satisfies those specific goals.

Yes, you can use these coding guidelines for both writing and reviewing code. They are applicable during either process to surface assumptions, keep changes surgical, and define verifiable success criteria.

You need to surface assumptions before an LLM writes code when facing complex requirements, utilizing a think-before-coding guideline to ask clarifying questions and state explicit assumptions before execution.