LLMs are already being used to generate code, but using them to generate and validate architecture is a fundamentally harder and more interesting problem. This talk introduces a practical approach to LLM-based design loops built on the BTABoK CLI and MCP (Model Context Protocol), where structured architecture artefacts — canvases, decisions, fitness functions — become the inputs and outputs of iterative AI-assisted design cycles. Rather than asking an LLM to freeform a system design, the loop grounds generation in BTABoK schemas, validates outputs against CoDL/CaDL constraints, and surfaces gaps for human review.
Drawing on BTABoK's Design concept — architecture as deliberate, constraint-aware shaping of solutions — the session is honest about where LLMs add genuine leverage (option generation, consistency checking, documentation) and where human judgement remains essential (trade-off resolution, stakeholder alignment, ethical constraints). Attendees leave with a concrete architecture for building their own design loop, not just a demo.