Why architecture matters here
MCP prompts fail when they're too flexible or too rigid. Too flexible → confusing UX. Too rigid → users can't adapt. The architecture matters because arguments + suggestions let the server guide safely.
With the pieces mapped, MCP prompts become guided workflows that host UIs can render richly.
The architecture: every piece explained
The top strip is publication. Server publish declares available prompts. Prompt template contains the content + argument placeholders. Arguments schema types inputs. Auto-complete returns argument suggestions.
The middle row is the flow. Client UI lets user pick + fill. Rendered prompt gets sent to model. Version + compat uses semver with deprecation. Localization supports i18n.
The lower rows are governance. Governance reviews prompts. Metrics track usage + edits. Ops handles deprecation, safety, and audit.
End-to-end flow
End-to-end: server publishes a "code review" prompt with args (file_path, focus). Client UI shows prompt library; user picks; auto-complete suggests file paths. User fills focus="security". Client renders + sends to model. Model responds. Metrics show usage.