karpathy-guidelines

To reduce LLM coding mistakes, enforce explicit assumptions and tradeoffs before implementation, ensuring the model surfaces hidden dependencies and defines a concise plan with verifiable success criteria.

Enforce surgical changes during LLM code review by requiring the model to limit unnecessary modifications and preserve existing logic, ensuring changes remain strictly scoped to the defined task.

Ensure LLM code generation follows a measurable plan by defining verifiable success criteria upfront and looping the execution until those measurable outcomes are explicitly verified.

LLMs make unnecessary changes during code refactoring because they lack explicit behavioral constraints, which can be mitigated by enforcing surgical rules that preserve existing code and limit scope creep.