What if we use the structured prompts from coding sessions, especially the ones which use arch design document, domain knowledge(UML, Statecharts what have you)

What if we use the structured prompts from coding sessions, especially the ones which use arch design document, domain knowledge(UML, Statecharts what have you), what team member to ask about X, for a large projects and fine tuned models. And these could all be made into tool calls for instruction following. Right now it seems teams manage a reasonably sophisticated LLM layer, MCPs and instruction following is one shot context window management dependent.