karpathy-guidelines

To stop an LLM from overcomplicating code during refactoring, apply behavioral guidelines that enforce simplicity first, demand surgical changes, and prevent speculative additions. This ensures the model writes only the minimum code necessary to solve the problem.

The best coding guidelines for LLM-assisted code generation enforce four core principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. This prevents common pitfalls like bloated abstractions, dead code, and incorrect assumptions.

To make an LLM ask clarifying questions before writing code, enforce a caution-biased guideline set requiring the model to think before coding. This approach prevents incorrect assumptions and ensures clear, goal-driven execution.

Yes, enforcing simplicity guidelines prevents LLMs from adding dead code by biasing the generation toward caution and surgical changes. The guidelines explicitly restrict speculative additions and bloated abstractions, ensuring cleaner, maintainable code.

To ensure an LLM matches existing code styles when making changes, enforce surgical changes as a core principle. This requires the model to execute goal-driven updates without introducing unnecessary complexity or unrelated features.

You should use LLM coding guidelines for software engineering tasks whenever you need to enforce caution and simplicity. Applying these guidelines to all coding tasks prevents overcomplication, bloated abstractions, and incorrect assumptions during code generation.