karpathy-guidelines

To prevent LLM coding mistakes during refactoring, enforce structured guidelines that surface assumptions, prioritize minimal surgical changes, and define verifiable success criteria before execution begins.

The best guidelines for LLM code review require listing explicit assumptions, defining 2-3 core guidelines, and creating a plan with verifiable checks to ensure outcome-driven execution before coding starts.

To stop LLMs from overcomplicating code changes, apply simplicity-first guidelines that prioritize minimal, well-scoped edits over broad speculative rewrites while preserving existing behavior.

You can enforce verifiable success criteria for LLM generated code by requiring planning steps that define measurable outcomes and tradeoffs upfront before any execution occurs.

Your LLM refactor introduces unexpected scope creep when guidelines fail to surface assumptions and tradeoffs upfront, leading to misinterpretation and broad speculative rewrites instead of surgical changes.

You should not use automated LLM coding guidelines for software engineering tasks that lack clear requirements or measurable success criteria, as the guidelines depend on upfront tradeoff analysis to function properly.