Mitigating AI-Generated Design Flaws: Shifting from Review to Contextual Guardrails?

I've been thinking about how AI-assisted system design could inadvertently increase our cognitive and technical debt if we're not careful. The recent discussions around RLHF prioritizing 'looks correct' rather than 'is correct' really hit home. We often talk about senior engineers reviewing AI-generated designs, but is that enough? Are we just moving the cognitive burden to a different stage? I'm wondering if a more effective strategy involves providing AI with richer, more structured contextual guardrails upfront, rather than solely relying on post-generation human review to catch subtle, yet critical, architectural flaws. For example, instead of just 'design a scalable microservice for user profiles,' what if we feed it specific non-functional requirements, existing architectural patterns, and even anti-patterns to avoid? Could this 'approval-seeking' nature of AI be better directed towards adhering to predefined architectural constraints, making it a design assistant that understands our architectural boundaries, rather than just a code generator? What approaches are others considering to embed architectural principles into AI's design process, beyond just prompt engineering?