“Generate UML from requirements” works best when the input reads like an architecture conversation: who uses the system, what entities exist, and what flows matter.
This checklist helps you get more reliable UML from an AI assistant (and fewer wrong diagrams).
1) Write requirements like system intent
Include:
- Primary users and roles
- Core entities (nouns) and relationships
- Two or three critical flows (happy path + one failure path)
- Any stateful lifecycle (order/task/request statuses)
- Constraints (auth boundaries, integrations, compliance, data retention)
2) Ask for the right views (not all views)
Start with a smallest-set that removes ambiguity:
- Class diagram: core domain model and ownership
- Sequence diagram: one critical end-to-end flow
- Activity diagram: workflow steps + branching
- State machine: lifecycle for stateful entities
Add more views only when you have a specific question you need to answer.
3) Validate before generating code
Look for:
- Duplicate or ambiguous naming (same concept, different words)
- Relationships without clear ownership (who “owns” the lifecycle?)
- Flows that skip state transitions (e.g., “Approved” without ever being “Pending”)
- Gaps between behavior and model (sequence diagrams referencing missing fields/classes)
4) Iterate in small loops
Small loops beat one giant prompt:
- Update description → regenerate views → validate → generate baseline → adjust
Next steps
- Read: AI Quickstart
- Try: Interactive demo