What you describe is a very different approach. It would require orders of magnitude more inference requests, but it would be missing out on all the power and "intelligence" of these new models because they wouldn't have sufficient context to make sensible decisions about what might be wrong or how to fix it. Also, there are not many hallucinations anymore now that these better models are available. But what you describe may work well, I'm not sure.