karpathy-guidelines

To prevent AI coding mistakes during code generation and refactoring, you must surface assumptions early, keep changes surgical to the requested scope, and define explicit verifiable success criteria before implementation.

AI introduces unnecessary complexity and speculative refactors when it silently chooses interpretations without explicit assumptions. Requiring minimal surgical edits limits changes to the requested scope and avoids added risk during software changes.

The best way to review code changes for hidden assumptions is to check for explicit uncertainties, verify minimal surgical edits, and ensure testable success criteria were defined before the software changes were implemented.

Yes, you can use surgical code edits for bug fixes without causing regressions by limiting edits to the requested scope, avoiding speculative refactors, and establishing verifiable success criteria to complete the task safely.

You need explicit assumptions for software refactoring whenever overengineering or hidden interpretations could cause regressions. Stating uncertainties and tradeoffs early prevents AI from silently choosing an interpretation during code changes.

To define verifiable success criteria for vague coding tasks, turn the implementation, review, and refactoring requirements into testable steps that explicitly confirm the requested scope was met without added complexity.