karpathy-guidelines

To stop LLM coding from generating bloated solutions, enforce simplicity-first implementation by discouraging speculative abstractions and unnecessary error handling. This approach limits edits to only what the task requires, preventing overconfident or unintended side effects.

The best way to ensure surgical changes during refactoring is to limit edits strictly to what is required and clean up only your own introduced artifacts. This prevents unintended side effects and maintains code integrity during LLM-assisted modifications.

Defining verifiable success criteria for ambiguous coding tasks involves turning vague requests into explicit, testable checkpoints. This goal-driven execution requires iterative verification to ensure outcomes match expectations without guessing.

LLM-generated code fails when requirements are ambiguous because it often makes hidden assumptions without explicit verification. Surfacing assumptions before implementation prevents overcomplicated solutions and ensures the code meets actual requirements.

Yes, you need to explicitly list assumptions before an AI refactors code to ensure clear thinking and minimal changes. Explicit assumption surfacing prevents misunderstandings and bloated solutions during the modification process.