AI Product Design Complete Guide: Generative UI, Trust, Feedback, Streaming, Agent UX, Korean Culture Specialization (2025)

Season 4 Ep 12 — If everything through Ep 11 was about "engineering reaching stability," Ep 12 is about being loved by users. No matter how good the model is, the product fails if UX is weak.

Prologue — "Good AI product ≠ Good UX"
Chapter 1 · Deterministic UI vs Generative UI
Chapter 2 · Designing for Trust
Chapter 3 · Feedback UX
Chapter 4 · Streaming, Latency, and Progress UX
Chapter 5 · Agent UX
Chapter 6 · Voice UX (Extending Ep 9)
Chapter 7 · Data and Learning UX
Chapter 8 · Onboarding and First Impression
Chapter 9 · Accessibility and Inclusion
Chapter 10 · Ethics and Responsibility
Chapter 11 · Korean Language and Culture Specialization
Chapter 12 · Five Real-World Cases
Chapter 13 · Ten Anti-Patterns
Chapter 14 · Checklist — Twelve Items Before Launching an AI Product
Chapter 15 · Next — Season 4 Ep 13 (Finale): "Business Models in the Generative AI Era"

Prologue — "Good AI product ≠ Good UX"

Many 2023-era AI products were just "ChatGPT wrappers": a text input, a response panel, and thumbs up/down. As of 2025, most products with only that format have either failed or survived only in narrow niches.

Why:

The ChatGPT mothership keeps getting stronger — simple wrappers have no differentiation
Users have started recognizing AI's weaknesses (hallucination, latency, sudden refusals) — trust design is required
You cannot demand "write good prompts" from non-technical users

So AI product design in 2025 is:

"UX design that builds trust within constraints."

This essay distills the concrete patterns.

Chapter 1 · Deterministic UI vs Generative UI

1.1 Deterministic UI

Buttons, forms, tabs, menus — results are predictable
Gentle learning curve, easy error correction
80 years of HCI optimization behind it

1.2 Generative UI

AI generates buttons, cards, charts, and forms on the fly
Works together with Tailwind / design-system tokens
UI morphs by context — extremely powerful but hard to predict

1.3 Hybrid patterns

Deterministic skeleton, generative content (lists, summaries, citations)
Render generated output as deterministic components (JSON to Card, Markdown to safe HTML)
Actionable Chip: the model proposes, the user taps, and a deterministic action fires

1.4 Examples

Notion AI's slash commands: deterministic + generative
Perplexity: generated summary and citations inside a deterministic layout
Cursor: deterministic IDE frame + inline generative code suggestions

"The freer the generation, the more deterministic guardrails you need."

Chapter 2 · Designing for Trust

2.1 Why trust is central

AI output is probabilistic. Every time, users silently ask "is this really right?"
One big error poisons even correct answers afterward
Trust is not accuracy. Trust is users being able to predict the AI's limits.

2.2 Seven patterns

Citation: show source documents, pages, and sections
Uncertainty display: language like "85% confident" or "insufficient info"
Editable: edit, delete, or regenerate a response in place
Visual distinction of origin: AI-generated areas marked with icon or background
Show work: expandable reasoning steps and search results
Freshness labeling: "Latest data: 2025-04-10"
Polite refusals: "I cannot help with that" + suggested alternative

2.3 What breaks trust

Confident answers with no sources
Hallucinations wrapped in "plausible" prose
Silent failure (no error message)
Excessive generative UI that reshapes the interface every time

Chapter 3 · Feedback UX

3.1 Collection

Thumbs up/down: instant feedback
Multiple-choice reasons (checkboxes): "inaccurate", "incomplete", "unsafe", "style mismatch"
Free text: rare but deeply insightful
The edited response itself is feedback (regenerate, edit)

3.2 Utilization

Downvote cases to regression-eval candidates
Upvote cases to preference training data (Ep 4 DPO)
Edit diffs to style/format learning

3.3 Placement

Inline buttons get 5 to 10x higher engagement than "go to dashboard" approaches
Immediately confirm "Thanks" after feedback
Be transparent about how personal data is used

Chapter 4 · Streaming, Latency, and Progress UX

4.1 Streaming vs batch

Streaming: feels fast, long responses stay engaging
Batch: complex formats (tables, JSON) or when the output must be validated as a whole

4.2 Handling latency

Skeleton UI: layout first
Typing indicator: "thinking..." or three dots
Progressive content: title, then summary, then body
Budget display: "about 5 seconds remaining"
Cancel and wait: user can abort

4.3 Progress state

Agent tasks: stages like "searching", "analyzing", "drafting summary"
Todo checklist: Claude Code style. Check off each completed step
Replay: review the process after completion

4.4 Preventing boredom

For long waits, use micro-interactions (animations, playful messages)
But excessive humor undermines professionalism

Chapter 5 · Agent UX

5.1 The three challenges

Progress visibility: what is it doing right now
Approval and abort: stop before risky actions
Recovery and replay: salvage even on failure

5.2 Progress visibility

Timeline: tool calls and results on a time axis
Todo: "1. Open file done 2. Editing... 3. Run tests pending"
Cost/Time meter: projected cost and time

5.3 Approval gates

Sensitive actions (payments, deletion, external sending) trigger a human-confirm popup
Shortcut: "auto-approve next time" option (explicit)
Prevent abuse of bulk approvals

5.4 Failure UX

Clear error reason + retry button
If the agent gives up on its own, present "reason and next alternative"
User can jump in via "need a hand?"

Share agent runs via link (internally)
Simulation and debugging
Onboarding material for new teammates

Chapter 6 · Voice UX (Extending Ep 9)

6.1 Design principles

Short, clear responses (summary of the Ep 9 12 principles)
Respect turn-taking (handle interruptions naturally)
Speak uncertainty ("what I understood is...")
Escalation trigger (after several failures, hand to a human)

6.2 Multimodal integration

Voice + screen: speak guidance, show cards and links on screen
Apple CarPlay / Android Auto integration
Smart home / wearables: no display means pure voice UX

6.3 Accessibility

Live captions and text alternatives by default
Speed and size controls
Adapt to quiet environments (auto-adjust TTS volume)

Chapter 7 · Data and Learning UX

7.1 Disclosing user data usage

State "where my data goes" in plain language
Whether it will be used for training, how long it is stored, how to delete it
Short privacy summary + detailed link

7.2 Feedback to learning loop

Set expectations about when user edits will be reflected in training
Distinguish personal model vs shared model

7.3 On/Off switches

"Do not use this conversation for training" toggle
Enterprise setting: organization-wide training opt-out

Chapter 8 · Onboarding and First Impression

8.1 The first 30 seconds

3 to 5 examples of "what you can do with this"
Clear data permissions and auth (OAuth, etc.)
"Sample question" buttons: one tap should work

8.2 Progressive disclosure

Do not expose every feature at once; teach along the way
New-feature tooltips, history-based recommendations

8.3 Error experience

The first failure is fatal — prepare recovery UX for the most common failures during onboarding
User education: explain "why it answered this way" interactively

Chapter 9 · Accessibility and Inclusion

9.1 Diverse abilities

Vision: screen readers, high contrast, font size
Hearing: captions, text alternatives, vibration
Motor: keyboard-only, voice-only
Cognitive: simple language, short responses, checklists

9.2 Language and culture

Multilingual support (UI + responses)
Local idioms, number formats, dates, currency
Cultural sensitivity (holidays, gender, greetings)

9.3 Economic access

Free / low-cost tiers
Offline mode (local LLM)
Low-spec device support

Chapter 10 · Ethics and Responsibility

10.1 Transparency

State clearly that AI authored the content
Mark the boundary between human and AI (especially in conversation)

10.2 Bias and fairness

Detect and correct outcomes unfavorable to specific groups
Test-set diversity (gender, age, region, occupation)

10.3 Labor and economy

Social responsibility toward jobs displaced by AI
Labor conditions for data labelers and evaluators

10.4 Environment

Energy consumption of large models and long contexts
Reduce via caching, routing, and smaller models

Chapter 11 · Korean Language and Culture Specialization

11.1 Tone and honorifics

Default to honorific form; allow soft mixing with informal depending on context
Consistency between the more formal style and the slightly softer polite style

11.2 Forms of address

"Customer", "User", "Mr./Ms. OO" — each brand has its own rules
Allow real-name and nickname settings

11.3 Cultural context

Holidays (Lunar New Year, Chuseok), public holidays, seasonal greetings
Sensitivity around age, school-year, and career questions
Caution when referencing occupation or hometown

11.4 Legal phrasing

Terms, privacy, marketing consent: clear and separated
Refusal phrasing for financial, medical, and legal advice

11.5 Korean mobile UX

Mixing Korean keyboard and voice input
Short-answer vs long-answer toggle
Balanced use of emoji and memes

Chapter 12 · Five Real-World Cases

12.1 AI coding products (Cursor-type)

Inline suggestions + chat + Todo list
One-key accept/reject of edits
Replay: full diff of what the agent did

12.2 Customer support (e.g. Zendesk AI)

Customer message and AI draft and agent edits
Show citations and knowledge-base links
"Auto-send" vs "edit-then-send" toggle

12.3 Writing (e.g. Grammarly, Notion AI)

Highlight-based edit suggestions
Style/tone sliders
"View original" / revert

12.4 Search and research (Perplexity-type)

Emphasize citations and sources
Auto-suggest follow-up questions
Collections and notes by topic

12.5 Agents (Manus, Devin-type)

Task board + Replay + external integrations
Approval on risky steps
Download/share deliverables

Chapter 13 · Ten Anti-Patterns

13.1 A lone prompt box

No onboarding, no samples, no suggestions. Blank-page terror.

13.2 No progress indicator

Users bail out of long agent runs.

13.3 Responses with no citations

Trust collapses.

13.4 Full confidence on hallucinated answers

False confidence is the worst UX.

13.5 Feedback collected, no sign of being used

Users give up quickly.

13.6 Swallowing errors silently

Unclear what failed and what to try next.

13.7 Automatic actions with no approval gate

A source of accidents.

13.8 Accessibility deprioritized

Always postponed, never shipped.

13.9 Missing AI disclosure

Regulatory violation + deceiving users.

13.10 Global tone ignoring Korean culture

Awkward in honorifics, address, and context.

Chapter 14 · Checklist — Twelve Items Before Launching an AI Product

Chapter 15 · Next — Season 4 Ep 13 (Finale): "Business Models in the Generative AI Era"

Technology, operations, and design are all covered. The final question is "how do you make money with this?"

Pricing models (Subscription / Usage / Hybrid / Seat / Outcome)
Cost structure and margins (API dependent vs in-house model)
GTM (B2C / B2B / Enterprise / Prosumer)
Building a data flywheel and moat
Getting past the "AI wrapper" critique
Regulatory and reputational risk as business impact
Investment landscape (VC, corporate, public)
M&A and integration
Korean startups going global
Season 4 retrospective
Season 5 preview (e.g. "Reinventing Data Engineering")

"No matter how strong the tech, without a business model it is a one-year product." The final Season 4 essay stitches tech, ops, and design together on the axis of money.

See you in the next one.

Summary: AI product design is engineering trust within constraints. The boundary between deterministic and generative UI, the 7-trust patterns of citation/uncertainty/edit, streaming/feedback/failure UX, the agent's progress/approval/replay, voice turn-taking, data transparency, 30-second onboarding, accessibility, ethics, and Korean-language/cultural fit. "A good AI product is a product that converses with users within constraints." The model is only the start; UX finishes the quality of the product.