karpathy-guidelines

To prevent overcomplicating LLM-assisted code changes during refactoring, enforce minimal, surgical edits that leave adjacent logic untouched. Surface assumptions explicitly, define testable success criteria, and require clear tradeoffs before considering any code modification complete.

Surfacing assumptions during AI-assisted code review involves explicitly calling out uncertainty instead of silently guessing. You must state missing context, evaluate tradeoffs, and define concrete verification checks to ensure the minimal code change is fully measurable and correct.

You verify success of minimal surgical code edits by defining concrete checks or tests before making changes. This ensures completion is measurable, prevents touching unrelated code, and confirms the simplest implementation path satisfies the original request without overengineering.

Your LLM changes unrelated code because it lacks enforced minimal scope constraints. By applying coding guidelines that require the smallest surgical edit and explicit tradeoff analysis, you prevent the LLM from overengineering simple fixes and modifying adjacent logic unnecessarily.

Yes, you can use coding guidelines to compare refactoring options before modifying a feature. This approach helps you choose the simplest implementation path, surface hidden assumptions early, and state concrete verification steps to confirm the change worked without overcomplicating the logic.