Every software product evolves after launch. But rarely does a SaaS user explore dozens of combinations of competing engines, interfaces, prompts, context strategies, agents, and integrations in search of the current winning strategy for every task. AI users do. Over 18 months of operating internal AI tooling inside a game studio, that behaviour started looking surprisingly familiar — and this talk is about what internal AI product teams can learn from LiveOps. Users discover effective combinations, share tactics, form strong preferences, and adapt whenever vendors change the underlying mechanics; effectively, they are building a meta. We watched teams settle into their own local versions of it. Some of it reflected measurable performance; some of it was folklore — model reputations, memorable successes, shared heuristics, prompt rituals. In a probabilistic system users cannot inspect, that is the only rational learning strategy available, and it drives real adoption and real spend. The comparison with LiveOps, however, is not one-to-one: a live team owns its own mechanics, and we do not. Vendors change models, defaults, prices, and packaging on their own schedule, and the patch notes arrive after the fact. Every release resets the meta — and so does every rule we introduce ourselves. We will look at how to read the meta your users are actually developing, how to tell measured performance from folklore, what happens to behaviour after every intervention, and how to recognise when yesterday’s rational decision has expired while developing an internal AI strategy.