AP3
All topics

New behavior / collecting interaction patterns before features

AI-Native Interaction

Design systems that can stream, use tools, recover, and express uncertainty without collapsing into a chat box.

Why this matters

AI-native interaction is an ownership problem as much as a model problem. A strong experience separates grounded knowledge from inference, keeps deterministic product actions deterministic, and makes interruption, cancellation, latency, and failure visible.

Concepts to learn

  • Streaming and realtime APIs
  • Tool calling and structured outputs
  • Grounding, retrieval, and memory
  • Interruption, cancellation, and latency masking
  • Confirmations and recoverable failure
  • Evaluation and product-quality criteria

Apply It in Casset

Casset is the working environment for this skill—not the curriculum itself.

  • Prototype Voice Liner Notes that distinguishes verified artist context, catalog metadata, model inference, and unknown information.
  • Build conversational catalog navigation where deterministic tools search and control playback; the model never owns playback state directly.
  • Explore an Artist Studio Assistant over real catalog and product data with explicit permission and provenance boundaries.
  • Test a Listening Companion that ducks or pauses predictably, streams quickly, accepts interruption, reflects voice state, resumes safely, and exposes errors.

Suggested project

Voice liner notes

A grounded, interruptible listening conversation that keeps source and playback ownership legible.

Open project notebook

Evidence of completion

These are completion conditions, not claimed achievements. Each stays planned until it links to a real artifact or verification.

  • implementationA streaming prototype with tool boundaries and visible stateplanned
  • testsEvaluation cases for grounding, unknowns, interruption, and recoveryplanned
  • observationA user session focused on trust and controlplanned
  • reflectionA note separating impressive model behavior from durable product behaviorplanned

Open questions

  • What should the model be allowed to propose, and what must remain a deterministic command?
  • How should uncertainty feel when the person is listening rather than reading?

Related notes

Books, talks, and useful things

  • article
    Designing AI Products

    Study interaction boundaries, not just model capability.

  • tool
    Structured outputs and tool calling

    Keep product state behind explicit tools.

Reflection journal

The model should expand the relationship with the music, not become a second product competing for attention.