The step that stands out is asking ChatGPT or Claude to compare your cheapest and most expensive ads. One thing I would add is to give it the same fields for every ad: spend, purchases, cost per purchase, audience and placement, and ask it to list what the winners share only where the sample is big enough. Otherwise it tends to explain noise with a confident story. I would also log each idea as a hypothesis, for example "UGC-style hook lowers cost per purchase", then check that against the linked new ad after a fixed spend, say a week or a set number of purchases. That turns the loop into a record of what you actually learned, not just a stream of new creatives.