I think it's the right approach, the cost of running these things as coding assistants is negligable compared to the benefit of even a slight model improvement.