karpathy-guidelines

LLM coding best practices involve stating assumptions, surfacing tradeoffs, and making surgical changes to modify only what is needed while keeping surrounding code stable. This prevents overcomplication by enforcing clear thinking before code generation.

Verifiable success criteria for LLM refactoring are defined by setting explicit, goal-driven outcomes before execution. You establish specific checkpoints to verify the refactor meets requirements, then iterate until the code passes those defined criteria.

To review generated code for unnecessary changes, apply guidelines that enforce surgical modifications. You verify that the LLM modified only what was needed for the task, keeping surrounding code stable and avoiding unrequested overcomplication.

You should use explicit guidelines for LLM coding tasks whenever your project requires precision and verifiable outcomes. They are essential during writing, reviewing, and refactoring to surface hidden assumptions and prevent behavioral mistakes.

Your LLM refactoring introduces unstable changes when it lacks guidelines to minimize modifications. Without rules enforcing surgical changes, the model overcomplicates the codebase by altering surrounding code instead of modifying only what is needed.