Here’s how I handle the AI customization tradeoff:
I split volatile vs stable: I keep business logic in code, prompts only for tone/format- I customize with data, not prompts: I use RAG/tools to inject context into one shared template- I treat prompts like code: versioning, evals, pruning one-offs

– I watch for prompt rot: model upgrades can break custom logic if I have no evald

I simplify the mechanism: I consolidate scattered customization into one system

Bottom line for me: customization isn’t the problem, untracked customization is.