karpathy-guidelines

To prevent LLM coding mistakes, enforce pre-implementation assumption surfacing and simplicity-first design principles. This approach requires explicit tradeoff evaluation before code writing, ensuring outputs remain minimal and free of speculative features or unrequested adjacent refactors.

AI generated code includes unnecessary refactoring because large language models frequently produce overcomplicated, misaligned outputs without strict guardrails. Applying surgical editing rules restricts code changes to only the requested scope, eliminating unplanned refactors of functional adjacent code.

The best way to enforce simplicity-first design during AI code review is to apply development guardrails that require minimal, non-speculative code. This ensures the AI solves only the requested problem without adding unrequested abstractions or excessive edge case error handling.

To stop AI from adding unrequested features, enforce surgical editing rules and simplicity-first design principles. This restricts modifications to only the user's explicit request, preventing the LLM from generating unplanned abstractions or performing unrelated code cleanups.

Yes, you can apply coding guardrails to refactoring tasks across all software engineering workflows. These guidelines ensure maintainable, minimal code modifications with clear verifiable success criteria, preventing unintended side effects in existing codebases.

You need pre-implementation assumption surfacing for development tasks before writing any code. This guardrail requires explicitly stating assumptions, tradeoffs, and alternative approaches to avoid misaligned LLM outputs and reduce technical debt.