Project Overview
I ran a Crazy 8s ideation sprint augmented by AI (ChatGPT) to explore mobile-first layouts for a “Shop Cards” experience, where Business cards ⊇ Personal cards (everything personal plus business-only benefits). I asked the AI to output XML/Markdown layout specifications for eight distinct patterns, then rendered and previewed them as interactive screens using code generated in-session.
The outcome: high-variance concepts, reusable UI primitives (e.g., TwoColumnCompare, BenefitMatrix, BottomDrawer), and a clear decision path to a hybrid solution.
ROLE & DURATION
Lead, AI Enablement | EPAM
Keywords: design thinking, Crazy 8s, AI-assisted ideation, mobile UX, layout specification, prototyping, prompt engineering
November 2025
Table of Contents
- Project Overview
- Scope disclaimer (Crazy 8s): What it is & where it fits
- Why this case study
- Framing the Problem
- Approach: Crazy 8s, augmented by AI
- Prompt structure
- The eight AI-generated directions (snapshot)
- From text to screens: turning XML into prototypes
- Visual Output in Chat GPT
- Selection criteria & decision
- What AI did well (and what needed human steering)
- What I’d do next
- Appendix: Crazy 8s in XML
Scope disclaimer (Crazy 8s): What it is & where it fits
Crazy 8s is a fast, divergent ideation exercise where each participant explores 8 distinct ideas in ~8 minutes to deliberately widen the solution space. It typically occurs after problem framing / research synthesis and before concept selection and prototyping i.e., in the early Ideate phase of the Double Diamond (between Discover/Define and Develop/Deliver).
In this case study, we intentionally did not optimize for brand communication, pixel‑perfect specs, or final tokens. Crazy 8s prioritizes breadth over polish. Our goal was to maximize idea diversity across components, compositions, and interaction patterns—and we achieved that with eight materially different, mobile‑first directions. Brand voice, legal/compliance copy, design tokens, and production‑level specs are deliberately parked for the convergence phase (prioritization, prototyping, validation).
Why this case study
Focus: process over brand. AMEX serves as a realistic scenario to ground requirements, but the study centers on how to use AI to run Crazy 8s and turn textual specs into visual prototypes quickly.
Outcomes: speed, breadth, and structure. AI amplified divergence (8 ideas in minutes), formalized design constraints inside prompts, and accelerated convergence via side-by-side previews.
Framing the Problem
Context (from Define > Ideate): Users come to shop for cards on mobile. Card types include Personal and Business, where Business = Personal benefits + business‑only tools. Early research and requirement mapping confirm confusion around where business benefits differ and how to compare quickly.
How Might We (HMW):
How might we clearly communicate that Business cards include all Personal benefits—plus business‑only tools—within a mobile card‑shopping experience that minimizes cognitive load, supports quick comparison, and remains accessible?
Constraints & signals
- Mobile-first, one‑glance comprehension; progressive disclosure for details.
- Comparison affordances that don’t overwhelm (legends, inline diffs, drawers).
- Strong accessibility (labels, contrast, text-only fallbacks).
- Content-resilient layouts that localize and scale across product lines.
Approach: Crazy 8s, augmented by AI
Traditional Crazy 8s: 8 ideas in ~8 minutes per participant.
AI-augmented variant I used:
- Codify constraints in the prompt (e.g., “business ⊇ personal,” mobile-first, a11y notes).
- Force representation as layout code (XML/Markdown) so every idea is visually describable and prototype-ready.
- Push for variety through explicit instructions (switch components, interaction patterns, disclosure modes).
- Immediate visualization: Convert XML to UI (React/HTML) inside ChatGPT’s code preview to get tappable prototypes.
- Rapid critique loop: Annotate pros/cons, adjust prompts, re-generate variants.
Prompt structure
Below is a simplified version of the prompt I used to kick off ideation.
- The prompt encodes the mental model and interaction palette.
- The output format becomes a shared contract between ideation and prototyping.
The eight AI-generated directions (snapshot)
Each was delivered as an XML-like layout spec. Here’s a condensed view of the eight with the core idea and signature components.
- Toggle + Stacked Benefits: Segmented control to switch Personal/Business; BenefitStack with legend (✓ = Personal, + = Business-only); sticky compare sheet.
- Side-by-Side Swipe: Split-pane carousel to swipe Personal vs Business; tap benefits for modal details.
- Progressive Quiz → Reveal: Two-step intent quiz; result panel that highlights “Everything in Personal, plus …” with expandable list and CTAs.
- Sticky Tabs + Diff Badges: Top tabs; BenefitMatrix shows deltas with “Business-only” badges; card lists per tab.
- Journey Narrative: Storyboard scenes (morning, month-end, team travel) mapping Personal vs Business benefits in real-life contexts.
- Chips Filter + Compare Drawer: Filter chips by benefit theme; mixed grid tiles; bottom drawer opens a compact comparison.
- AR Card Stack: Flip/long-press interactions on 3D card stack; explicit a11y text comparison fallback.
- Truth Banner + Accordions: One-line truth banner; accordions for Personal and Business-only extras; inline micro-compare rows.
Design principle reinforced across variants: Communicate the set‑inclusion relationship (Business = Personal + extras) via legends, badges, banners, and row-level diffs.
From text to screens: turning XML into prototypes
One advantage of generating structured layout specs (XML worked better for me) is that they’re trivially compilable into previewable screens.
Workflow I used
- Generate XML in ChatGPT.
- Transform to JSX (small script or in-chat conversion) while preserving component names (e.g.,
TwoColumnCompare,BottomDrawer). - Preview directly in the chat code runner or a local sandbox to check spacing, hierarchy, and thumb reach.
Benefits
- Removes handoff friction between ideation and UI.
- Elevates evaluability: you critique interactions, not just wire boxes.
- Encourages a design system mindset (reusable primitives across variants).
Visual Output in Chat GPT
Selection criteria & decision
I scored concepts using a light rubric (1–5):
- Clarity of the “Business ⊇ Personal” model
- Scanability on mobile (1-screen comprehension)
- Effort to implement (tokens, components, backend tie-ins)
- A11y strength (labels, fallbacks)
- Extensibility (localization, more product lines)
Shortlist chosen for convergence
- #1 Toggle + Stacked Benefits as the primary (fast comprehension, clean compare sheet, low build risk).
- #6 Chips + Compare Drawer as a power-user layer (select by benefit themes, bottom-drawer compare).
These combine into a hybrid: clear entry, progressive disclosure, and flexible exploration without kicking users out of flow.
Metrics & leading indicators
- Comprehension time: Can users articulate “Business ⊇ Personal” within 3–5 seconds?
- Tap intent mix: % opening the compare sheet vs. card details indicates whether the model is understood.
- Scroll depth: Are users discovering business-only extras without pogo-sticking?
- Quiz completion rate (for the #3 variant, if used).
What AI did well (and what needed human steering)
Strengths
- Volume & variety on demand (8+ patterns without creative stall).
- Structure: predictable component schemas that are easy to compile.
- Consistency: a11y reminders, legends, and repeated primitives.
Where human judgment mattered
- Picking the right trade-offs between clarity and novelty.
- Copy tone and market-specific terminology.
- Polish: spacing, thumb zones, and device-specific micro-interactions.
- Compliance nuance (e.g., disclosures, eligibility messaging).
What I’d do next
- Usability micro-test: measure comprehension and path choices.
- Refine the hybrid (#1 + #6) with guardrails for localization and variable benefit sets.
- Design tokens: tighten type scale, grid, and spacing so components travel well across journeys.
- Document the pattern as a shareable recipe (prompt + components + a11y notes).
Crazy 8s with AI is not about replacing creative judgment, it’s about removing friction between ideation and evaluable UI. By insisting on layout-structured output and instant code previews, we moved from “ideas on paper” to tappable, testable screens in one sitting. The constraints let us validate clarity across multiple interaction modes and extract reusable design patterns for future journeys.
Appendix: Crazy 8s in XML
(Output from the prompt)
1) Toggle + Stacked Benefits with “Business ⊇ Personal” Callouts
2) Side-by-Side Swipe (Split Pane Carousel)
3) Progressive Quiz → Personalized Reveal
4) Sticky Tabs + Inline Diff Badges
5) Timeline / Journey Story (Narrative Comparison)
6) Chips Filter + Inline Compare Drawer
7) AR Card Stack (3D Cards + Peekable Benefits)
8) “One Line Truth” Banner + Inline Accordions