Cool idea, hadn't considered the use of lesser models picking up the implementation set by the strong model plan.
Question regarding the actual code generated, is this a priority to you beyond looking at it meets the design doc? For example, if the LLM goes ahead and builds the feature that passes the tests, but the code contains duplicate functions, abstractions etc, would you steer the LLM to fix that, or even dive in yourself to rearchitect?
The same applies here and in any other tech field, nobody ever regrets that they learned "too much".
Pick your poison and learn everything you can!