| name | dec-ai-native-patterns |
| description | AI-native design engineering patterns — generative/adaptive UI, streaming as interaction, prompt-as-interface, human-in-the-loop / agentic UX, eval-driven development, designing for uncertainty, latency and cost as design constraints. Use when designing any AI/LLM-powered feature, building agentic or chat UX, handling streaming/probabilistic output, or planning for model error and cost. |
| tags | ["ai-native","llm","agentic-ux","streaming","evals","always-on"] |
AI-Native Design Engineering
The new layer, increasingly central to the role. Probabilistic systems need UX that plans for being wrong.
- Generative / Adaptive UI: Interfaces assembled or personalized at runtime by a model rather than fully predetermined.
- Streaming as Interaction: Token-by-token output, partial results, and stop/regenerate controls are core interaction patterns, not edge cases.
- Prompt-as-Interface: Natural language as an input modality alongside (not replacing) direct manipulation — with affordances that teach users what the system can do.
- Human-in-the-Loop / Agentic UX: Design review, approval, correction, and undo around autonomous actions; surface confidence and provenance.
- Eval-Driven Development: Treat model behavior as a testable surface — eval sets and regression checks for AI features, analogous to visual regression for UI.
- Designing for Uncertainty: Probabilistic outputs need graceful failure, transparent limitations, and easy escape hatches. The system will be wrong; the UX plans for it.
- Latency & Cost as Design Constraints: Model choice, caching, and streaming are UX decisions, not just infra ones.
How to apply
- Make streaming first-class: show partial output immediately, always offer stop and regenerate.
- For agentic actions, design the review/approve/undo loop before the autonomous capability — surface confidence and provenance.
- Teach the user what the system can do via input affordances, not a blank prompt box.
- Build an eval set early and run it like a regression suite; "looks good once" is not a test.
- Plan the wrong-answer path (graceful failure + escape hatch) as carefully as the happy path.
- Treat latency/cost/model choice as UX levers — cache, stream, and pick the smallest model that meets the bar.
