What it gives. Rules for how an AI product behaves. Treats the context window as a design material (token budgets, summarize vs retrieve, graceful degradation), shapes conversation (turn length, repair sequences, grounding), decides when the model should render UI vs plain text, staggers capability through progressive disclosure, and defines when the agent leads vs the user.
How to fire it.
Design the turn-taking and repair flow for a booking assistant, and when it should render UI vs text.
Payoff. The product reads as a designed experience, not a chat wall.
One of the six skills in Owl-Listener/ai-design-skills
. Covers the thread's Layer-4 concepts: context-window design, conversation patterns, generative UI, progressive disclosure, multimodal orchestration, mixed-initiative flow, frustration detection, feedback loops.