karpathy-guidelines

To prevent LLM coding mistakes during code review and refactoring, enforce behavioral discipline by requiring explicit assumptions, simplicity, and surgical edits. This ensures the assistant makes only necessary changes and states assumptions instead of guessing.

The best way to make surgical edits when fixing bugs with an LLM is to restrain modifications to only what the request requires and clean up only the resulting artifacts. This prevents unnecessary abstractions and unrequested configurability.

To convert refactoring tasks into verifiable success criteria, translate the requirements into measurable goals, often via tests. You then loop the execution until those goals are met to ensure the code works as intended.

Your LLM writes unnecessary abstractions when refactoring code because it lacks enforced simplicity constraints. By applying guidelines that bias toward the minimum code needed to solve the stated problem, you prevent speculative and unrequested configurability.

Applying engineering guidelines does help with ambiguous code requirements by forcing the LLM to think before coding. It surfaces assumptions, clarifies ambiguous requirements, and asks when multiple interpretations exist.

You should not skip an assumption audit for code generation when requirements are ambiguous or complex. Forcing explicit assumptions and a concrete step-by-step plan is critical to reducing speculative edits and ensuring verifiable success.