karpathy-guidelines

To prevent LLM coding mistakes, enforce assumption surfacing, simplicity, and verification-driven execution. This ensures changes remain accurate, minimal, and testable by requiring explicit success criteria and goal-to-test planning before implementation.

You can stop AI overengineering and unrelated refactoring by enforcing surgical edits and scope control. This limits modifications to only what the task requires, actively preventing style drift and speculative feature additions.

To clarify ambiguous requirements before LLM code generation, force the AI to surface explicit assumptions and ask for clarifications. This think-before-coding approach turns vague requests into verifiable goals with structured plans.

Yes, this approach works for writing new code and reviewing existing code. It applies to coding, reviewing, and refactoring tasks, ensuring every changed line is traceable to the original request through verification-driven execution.

The best way to ensure AI-generated code is testable and minimal is to apply goal-driven execution. This method transforms vague requests into verifiable goals with explicit success criteria, enforcing minimal necessary complexity.