The signal I’d care about is whether they can scope the task, prompt the model, review the output, explain tradeoffs and clean up what the model got wrong.