The maintenance tax is real. Custom prompts and workflows rot fast because models keep changing underneath them. I have seen teams do better when they ship strong defaults and treat customization as an escape hatch, not the main path. Prompts need versioning, evals, and rollback like any code. Customization usually pays off in data, retrieval, and domain workflows. It rarely pays off in per customer prompt forks. Most teams end up simplifying their AI stack over time, not expanding it.