karpathy-guidelines

Prevent LLM coding mistakes during refactoring by enforcing surgical changes that modify only what is needed, surfacing hidden assumptions explicitly, and defining verifiable success criteria before implementation.

Surgical changes in LLM code generation mean modifying only what is needed to fix an issue or add a feature, intentionally preserving existing behavior to avoid overcomplication and reduce unintended side effects.

Surface hidden assumptions when using an LLM to write code by requiring explicit assumption statements before implementation, pushing back on unnecessary changes, and defining measurable success criteria.

Yes, these guidelines apply during writing, reviewing, or refactoring code, providing behavioral rules to surface tradeoffs, push back when warranted, and define verifiable success criteria across all stages.

Define verifiable success criteria for LLM coding tasks by establishing a clear plan with measurable checks before implementation, ensuring goal-driven execution and avoiding overcomplication.

LLMs make unnecessary changes when refactoring because they lack behavioral guidelines for surgical changes, leading to overcomplication. Applying explicit tradeoff surfacing and goal-driven execution prevents this.