AI / Conversational UX2025

Mini Artifacts

An experiment in giving AI a spatial layer — can conversation-aware cards that spawn, evolve, and dissolve reduce cognitive load in planning tasks?

6 demo scenariosDual-card architectureResearch prototype
React 19TypeScriptDesign ResearchConversational AI

Hero image — placeholder

TL;DR

Built an interactive research prototype that adds a spatial card layer to AI chat — living cards that persist and mutate, ephemeral cards that appear and dissolve — testing whether AI-managed visual context reduces cognitive load in multi-step planning.

Overview

Role

Design Researcher & Prototyper

Domain

AI / Conversational UX

Stack

React 19, TypeScript, Tailwind, Vite

01 — Observation

The linear stream problem

I noticed a pattern across ChatGPT, Claude, and Gemini: when users plan complex, multi-step tasks through chat, structured information — checklists, budgets, timelines — gets buried in the message stream.

🧠

Cognitive overload

Users juggle structured data and conversational context simultaneously in working memory.

📜

Lost context

That checklist from 20 messages ago? Gone. Users scroll back repeatedly to find structured information.

No persistence

Existing artifacts are user-initiated snapshots. Nothing evolves automatically as the conversation progresses.

The gap: AI has no mechanism to proactively manage structured, evolving visual context tied to the conversation lifecycle.

02 — Hypothesis

A spatial layer for AI

What if we gave the AI a second output channel? Not just text in a stream, but cards it can spawn, mutate, and dismiss — living alongside the chat in a three-column layout.

CardsChecklist, Budget
ChatConversation stream
CardsTimeline, Tips

AI-initiated

The AI decides when a card reduces cognitive load. Users never request them — they just appear when needed.

Evolving state

Cards mutate in-place as the conversation progresses. A checklist gains items. A budget fills up. Version-tracked.

Automatic lifecycle

Contextual cards dissolve when the topic shifts. Persistent cards stay. The workspace stays focused automatically.

03 — Experiment

Two modes, one system

The key insight during prototyping: not all context has the same lifespan. A budget tracker should persist and evolve. A definition should appear, help, and get out of the way. This led to a dual-card architecture.

Living Cards

Persist & evolve

Spawned once, mutated in-place across the entire conversation. Version counter tracks changes. Update flash draws attention to mutations.

Checklists with items completing in real-time
Budget trackers with category breakdowns
Timelines filling day-by-day
Ephemeral Cards

Appear & dissolve

Contextual cards for a specific moment. Auto-dismissed when the topic shifts. Tips, definitions, references that serve their purpose and fade.

Definitions when new terms appear
Cultural tips during travel planning
Quick references that don't need to persist

04 — Test Scenarios

Six scenarios

To test whether the pattern generalizes beyond a single use case, I built six scripted demos across different domains. Each spawns the same card types with domain-specific content.

🗾

Japan Trip

5-day itinerary on a $3,000 budget. Flights, accommodation, day-by-day activities.

TravelBudgetItinerary
🍽️

Dinner Party

Host 8 guests on $200. Menu, shopping, cooking timeline, day-of execution.

FoodPlanningHosting

Freelance Sprint

Scope a 5-day web dev project. Deliverables, invoicing, daily sprint plan.

WorkScopingInvoicing
🚀

Startup Launch

Launch a SaaS product in 5 days with $5K marketing budget.

StartupMarketingGTM
🏠

Kitchen Reno

Renovate a 120 sq ft kitchen on $15,000. Permits, contractors, 5-week schedule.

HomeRenovationScheduling
🎸

Guitar Learning

Learn guitar from zero in 5 weeks. Gear, skill milestones, weekly practice plan.

LearningMusicPractice

Same card system, six different domains. If the pattern only worked for trip planning, it would be a feature. Working across all six suggests it's a generalizable interaction model.

05 — Design Rationale

Why these choices

Each design decision in the prototype was made to isolate a specific variable in the experiment.

Staggered spawn timing

200ms stagger between cards. Testing whether choreographed entry reduces perceived complexity vs. simultaneous spawn.

Update flash system

1.4s glow when a card mutates. Testing whether peripheral attention cues are enough to register state changes without reading.

Viewport-locked layout

Cards and controls never leave view. Testing the hypothesis that persistent visibility eliminates context-switching scrolls.

Color-coded categories

Per-demo color systems for tags and budget categories. Testing whether color-scanning outperforms text-scanning for structured data.

06 — Validation

How to prove it

A prototype proves feasibility, not value. This is the evaluation framework I'd use to determine whether the spatial card model actually helps — or just looks interesting.

Core hypothesis: AI-managed spatial cards reduce cognitive load and improve task outcomes in multi-step planning conversations, compared to a standard linear chat.

A/B test structure

Condition A

With cards

Three-column layout. Living cards + ephemeral context cards alongside chat.

Condition B

Chat only

Standard single-column chat. Same AI responses, same information — all inline.

What to measure

Quantitative

Task comprehension

Post-task quiz: "What's left on the checklist?" "How much budget remains?" Measures information retention.

Higher accuracy = valid
Quantitative

Scroll distance

How far users scroll back to find earlier information. Cards should reduce re-scrolling significantly.

Less scrolling = valid
Quantitative

Task completion quality

Rate the final plan output (completeness, budget accuracy, timeline coverage).

Better plans = valid
Qualitative

NASA-TLX workload

Standard cognitive load survey after each condition. Measures mental demand, effort, and frustration.

Lower load = valid

Kill criteria

Cards are ignored

Users don't look at side panels and rely entirely on the chat stream.

Comprehension doesn't improve

Post-task quiz scores are equal or worse than chat-only.

Users want control

Users consistently want to dismiss, rearrange, or override AI card decisions.

07 — Status

Where this stands

The prototype is built and functional across six domains. It proves the interaction model is technically viable and generalizable. What it hasn't proven yet: whether users actually benefit from it. The next step is running the validation framework above with real participants to find out if this is a useful pattern — or just an interesting one.