karpathy-guidelines

To apply behavioral guardrails against LLM coding mistakes, state assumptions explicitly, prioritize minimal design, make surgical edits, and define verifiable success criteria before coding. This ensures clear, minimal, and verifiable changes across coding tasks.

The best way to keep LLM code refactoring surgical is to modify only what is necessary and avoid touching unrelated code. Prioritizing simplicity and avoiding unnecessary abstractions ensures minimal design scope during project refactors.

Yes, you need to define verifiable success criteria before generating code with an LLM. Goal-driven execution requires stating explicit assumptions and looping until the verification criteria are met to ensure safe, verifiable changes.

You can use these code review guardrails for trivial tasks, but the guidelines bias toward caution over speed. For trivial coding tasks, use your own judgment to decide if applying behavioral guardrails is necessary.

Your LLM generates unnecessary abstractions because it lacks behavioral guardrails prioritizing simplicity first. By enforcing minimal design and explicit assumptions, you prevent the LLM from creating complex code structures during software engineering tasks.