karpathy-guidelines

To stop LLM coding from overcomplicating your codebase, apply behavioral guidelines that enforce surgical changes and emphasize minimum code to solve the problem without speculative features. This approach targets code reviews and refactoring to systematically improve code quality.

The best way to define verifiable success criteria for LLM code generation is to establish goal-driven execution loops that continue until the objective is verified. This ensures LLM coding targets are met without unnecessary modifications or overcomplication.

To make surgical changes during code refactoring with an LLM, follow guidelines that restrict modifications to only necessary parts of the code and require cleaning up any resulting mess. This minimizes risk and maintains overall code quality.

Yes, you can use these code quality guidelines for standard manual code reviews. The behavioral rules focusing on simplicity and surgical changes apply universally to refactoring and code quality improvement, regardless of whether an LLM generated the code.

LLM generated code often includes unnecessary speculative features because models lack inherent behavioral constraints to enforce simplicity. Applying specific coding guidelines corrects this by prioritizing minimum code to solve the problem and avoiding overcomplication.