The “Who Gets The Aux” Solution
Platform: Conversational AI concept / Mobile
Context: Solo design exploration - Multi-turn intent refinement for multi-user agentic personalization
Tools: Figma (screen mockups), Notion (dialogue design and persona documentation)
References: Screenwriting dialogue principles; CDI Foundations coursework
The Problem
I've lived in shared spaces my whole life, family homes, share houses, roommates, partners. And I've noticed that the small things matter more than they should. Who controls the music is never really about the music. It's about whether your presence in the space is acknowledged.
This is a conversational design exploration of what it would look like to build an AI mediator for shared listening, a system that makes the negotiation invisible so two people can just listen.
My Role
This was a solo conversational design exploration. My focus was on:
- Defining the system's persona and behavioral constraints before writing a single line of copy
- Designing multi-turn intent loops and system utterances—including system reflections, rebalance prompts, and conflict-resolution logic.
- Designing tone chip copy that describes user experience rather than system mechanics
- Documenting the iteration rationale behind specific copy choices
System Persona Brief
Before writing a single utterance, the persona is defined. Every word the system says is testable against this brief.
How it Works (The 3-Step Flow)
The system works in three phases. The design challenge is different at each stage.
Phase 2: The Vibe Check: Tone Chip Copy
The tone chips replace sliders and percentage inputs with plain human language. The design constraint: each sub-label must describe what the user's experience will be, not what the algorithm will do.
The framing question
The question that introduces the chips matters as much as the chips themselves.Why it changed: By reframing the prompt from control to collaboration, the system minimizes the "social tax" of negotiation. It shifts the user mental model from algorithmic manipulation to co-creation, optimizing for perceived fairness before the playlist is even generated.
Vibe Check: Annotated Dialogue
Scenario A: Two users, different chips
Alex picks 'Drive it'.
Mia picks 'Surprise me'.
The system names the dynamic and explains what it did.
Scenario B: Both users pick 'Drive it'
Adversarial Edge Case: Multi-user intent conflict. When both users attempt to dominate the session logic, the system relies on predictive mediation rules. Instead of picking an arbitrary "winner," the system executes a character beat, acknowledging the shared competitive intent, reflecting the tension transparently, and dynamically weaving both profiles into a high-contrast blend.
Phase 3 - One-Tap Rebalance Prompts
These ghost buttons let users adjust the playlist in one tap without starting over. Design principle: each prompt describes the desired experience, not the technical action.Iteration: Rebalance prompt tone
Why it changed: The first set is a settings menu. The second set is what you'd actually say to a friend who was DJing. The persona is a cool DJ, the copy has to earn that framing at the word level, not just claim it in the case study write-up.
Rebalance prompts in sequence
This shows how the prompts work as a conversation, not just one-off buttons.
The Screenwriter's Lens
My screenwriting background shapes how I approach conversational copy. Three principles from that practice apply directly here. I treat system reflections as character beats. In screenwriting, subtext is everything. In multi-user AI systems, the subtext is often interpersonal friction. My design logic uses the system's voice to diffuse that friction by mirroring user intent with absolute transparency.
Reflection
The hardest part of this design wasn't the conflict scenarios, it was the neutral state. When both users are happy and the playlist is running, the system has nothing to do. Getting the reflection line right (specific enough to be useful, short enough to stay out of the way) took the most iteration.
The Social Tax insight held up through every scenario: the system's job isn't to be a better recommendation engine. It's to make the negotiation invisible, so two people can just listen.
What I'd do next
- User testing the tone chips with real collaborative listeners, particularly whether 'Surprise me' reads as genuine enthusiasm or polite surrender
- Designing the edge case where three or more users are in the session as the conflict logic gets significantly more complex
- Testing whether the one-tap rebalance prompts reduce or increase social friction. There's a version where offering 'More of Mia's taste' as a button makes the negotiation more visible, not less