Start with clients, edge/API layer, application services, primary data stores, caches, queues, and external dependencies only where required. Label important flows and identify the source of truth for each key entity.

Then choose two or three high-risk areas for deeper analysis: a hot read path, a consistency-sensitive write, a fan-out problem, a large data store, or a failure-prone integration. Explain the baseline, estimate the bottleneck, and evolve the design only to address a demonstrated constraint.

Use diagrams as communication aids, not decoration. Keep names consistent, show asynchronous boundaries, mark ownership, and distinguish synchronous calls from events. If you add a component, explain which requirement it satisfies and what new failure mode it introduces.