karpathy-guidelines

To reduce LLM coding mistakes, enforce clear assumptions, simplicity, and surgical refactors that limit edits to the user's request. Define verifiable success criteria with test-backed goal loops to verify progress and prevent accidental collateral changes.

Prevent overengineering in AI-generated code by applying simplicity-first guidelines that discourage unnecessary abstractions and speculative error handling. Surface ambiguous interpretations explicitly rather than picking them silently during code review.

Ensure surgical changes during LLM refactoring by tying edits directly to user requests and avoiding risky refactors. This approach preserves unrelated behavior and prevents accidental collateral changes from occurring.

Unintended collateral behavior changes occur when LLMs lack enforced surgical edits and clear assumptions. Apply guidelines that tie edits directly to user requests and establish verifiable success criteria with test-backed loops to prevent this.

Yes, behavior guidelines turn coding tasks into testable success criteria by establishing a goal-driven execution loop. This loop verifies progress and defines exactly what tests or checks prove the code result is correct.