karpathy-guidelines

To stop LLMs from over-engineering code, apply behavioral guardrails that enforce a simplicity-first approach and surgical editing. This prevents premature abstraction and ensures changes are limited strictly to the requested task, preserving existing style and preventing noisy diffs.

Surgical editing in LLM coding is the practice of limiting code changes strictly to the requested task to prevent bloated code. It preserves existing style and prevents noisy diffs by avoiding unnecessary refactoring and speculative feature creep.

Goal-driven verification transforms vague coding requests into verifiable success criteria, ensuring code is tested and functional before completion. This approach requires strict adherence to explicit assumption surfacing to validate the final implementation.

Yes, you can apply these coding guardrails to both feature implementation and code reviews. They are designed for all software development tasks, including bug fixing, by enforcing simplicity, surgical precision, and goal-driven verification across the entire workflow.

LLMs make silent assumptions during software development tasks due to a lack of enforced behavioral constraints. Requiring strict adherence to explicit assumption surfacing forces the model to state its assumptions, preventing unexpected integration issues and bloated code.