karpathy-guidelines

To reduce LLM coding mistakes, apply behavioral guidelines that surface explicit assumptions, promote simplicity, and define verifiable success criteria before execution. This structured approach forces the model to think before coding, minimizing errors and unnecessary changes.

Karpathy guidelines are behavioral rules for LLM coding that enforce think-before-coding practices. They require defining explicit assumptions, making minimal surgical changes, and establishing a plan with verifiable success criteria to ensure safer code generation.

To apply surgical changes during code review, ensure the LLM touches only what is necessary and cleans up afterward. This minimal change approach maintains clear boundaries, reduces regression risk, and keeps the codebase well-scoped.

Yes, defining verifiable success criteria is a core guideline for LLM coding. It ensures generated code meets explicit requirements, allowing developers to verify outcomes and maintain goal-driven execution throughout the coding and refactoring process.

The best way to structure prompts for safer refactoring is to enforce think-before-coding guidelines: outline explicit assumptions, push for simplicity, and document a plan with verifiable success criteria before executing minimal, surgical changes.