I think the one thing the author is underestimating, especially in his "first pillar" is that he is able to coach the models into great results because he already knows the lay of the land. I often make the same mistake. I see people struggle with GenAI, and feel flabbergasted how they succeed to fail, but then if you observe how they work the tool, it is clear they have no idea what to ask or how to evaluate and iterate on the responses.