karpathy-guidelines

To prevent LLM coding mistakes during refactoring, apply behavioral guidelines that enforce surgical edits, ensuring the AI modifies only necessary code without introducing speculative features or unnecessary complexity.

Coding guidelines for AI-assisted software development should focus on explicit assumption surfacing, minimizing code complexity, performing surgical edits, and establishing verifiable success criteria to mitigate common large language model errors.

Verifiable success criteria for AI code generation tasks are defined by establishing clear, testable goals before implementation, ensuring that code changes remain minimal, well-justified, and directly address the specified requirements.

AI-generated code often introduces unnecessary complexity because models tend to overcomplicate solutions; applying focused guidelines encourages writing the minimum necessary code and explicitly surfacing tradeoffs to reduce bloat.

Yes, you can use these coding guidelines for reviewing AI agent code changes, as they provide actionable principles to surface hidden assumptions, verify success criteria, and ensure modifications are surgical and minimal.

You should be cautious with automated refactoring when tasks lack clear verifiable success criteria or when the AI risks modifying unnecessary parts of the codebase, as guidelines emphasize surgical, minimal changes only.