karpathy-guidelines

To prevent common LLM coding mistakes, apply behavioral guidelines that enforce explicit assumptions, surface ambiguities, and prompt the model to present multiple interpretations before writing code. This reduces unintended side effects and ensures safer code generation.

The best way to ensure surgical changes is to instruct the LLM to edit only what is necessary, preserve existing code style, and use minimal, purpose-built modifications without introducing unnecessary abstractions or unrelated edits.

To get an LLM to think before coding, use rule-based prompts that require it to explicitly state assumptions, surface ambiguities, and choose the simplest verifiable approach before generating any code blocks or refactoring solutions.

Your LLM introduces unnecessary abstractions because it lacks strict simplicity constraints. Enforce guidelines that demand minimal, purpose-built changes and surgical edits to prevent the model from over-engineering solutions beyond the requested scope.

Yes, you can use these coding guidelines for writing new code and reviewing existing code. The heuristics apply across refactoring and generation workflows to surface tradeoffs, define verifiable success criteria, and encourage safe changes.

To define verifiable success criteria for LLM-generated code, establish goal-driven execution rules that require the model to loop with checks until the output is verified. This ensures the generated code meets explicit functional requirements.