Case Study

How Gen AI
is rewriting
the creative playbook

Trip Atlas was built at the intersection of artificial intelligence and travel discovery — a live experiment in what happens when you stop thinking in suites and start thinking in agents.

The Disruption

Every generation gets
its defining disruption.
This one is different.

Print democratized publishing. The internet eliminated geography. But Generative AI is doing something more fundamental: it's collapsing the distance between intention and execution.

Previous disruptions changed how work was distributed. Gen AI is changing what work means — who does it, how fast, and what "good enough" looks like. For creatives, this isn't a productivity upgrade. It's a renegotiation of the creative contract.

1450s
The Printing Press
Democratized text. Threatened scribes. Created publishing. Took decades to absorb.
1990s
The Internet
Eliminated distribution costs. Disrupted every media model. Took a generation to restructure industries.
2010s
Mobile + Cloud
Untethered the workplace. Created the app economy. Collaboration became ambient.
Now
Generative AI
Collapses the gap between idea and output. Changes not just how work is done — but who can do it and why.
The Creative Shift

Speed. Information. Access.
All three, at once.

"We used to measure creative output in days. Now we measure in prompts. The bottleneck is no longer production — it's judgment."

The new creative economy
01

Speed without sacrifice

What once required a team, a brief, and two weeks of iteration now emerges in hours. Concepts are tested, refined, discarded — faster than any agency sprint. The creative process isn't slower because of perfectionism; it's liberated from it.

02

Information on demand

Research, synthesis, competitive analysis, tone guidance — tools now surface what used to live in expensive consultants or months of market research. A solo builder today commands resources that once required an enterprise.

03

Radical democratization

The barrier to entry for sophisticated creative and technical work has collapsed. A single person with the right set of AI tools can ship what previously demanded a full-stack team — designers, copywriters, engineers, strategists.

04

Judgment becomes the craft

The new creative skill isn't execution — it's curation. Knowing which output is right, which tool to trust, which prompt opens the door. The creative director has become the primary role, even for solo operators.

From Suite to Agents

The end of the
integrated suite.

For decades, software worked in ecosystems: Adobe Creative Suite, Microsoft Office, Google Workspace. One vendor, one login, one philosophy. Files flowed from app to app by design.

Gen AI has broken this. The new model is a constellation of best-in-class agents — each with its own strengths, personality, pricing, and interface. No single tool dominates. The creative's job is now orchestration.

Before · The Suite Era
Integrated. Closed. Loyal.
  • One vendor owned your entire workflow
  • Files designed to flow between products
  • Switching costs kept you locked in
  • Features determined by roadmaps, not needs
  • Learning curve shared across the suite
  • One monthly subscription covered everything
vs
Now · The Agent Era
Specialized. Open. Fluid.
  • Best-in-class tool for each task
  • Outputs move via clipboard, API, or prompt
  • Switching is frictionless — often free to try
  • Tools evolve weekly, not quarterly
  • Each tool has its own tone, logic, and "flavor"
  • Multiple subscriptions, composed to the job
The Agent Landscape

Every tool has its
own voice.

The modern creative stack isn't a suite — it's a cast of characters. Each agent brings a distinct capability, flavor, and philosophy. Knowing which one to reach for, and when, is the new creative literacy.

Writing & Strategy
Claude / ChatGPT
Long-form reasoning, tone calibration, strategic framing. Best for thinking through problems, not just generating answers.
reasoning copy strategy
Visual Generation
Midjourney / Firefly
Concept art, moodboards, hero imagery. Each model has a distinct aesthetic signature — Midjourney's painterly vs Firefly's commercial-clean.
imagery brand concepting
Code & Build
Cursor / v0 / Lovable
From UI component to full-stack app. Scaffolding in minutes, iteration in seconds. The line between design and development is dissolving.
frontend backend prototyping
Research & Synthesis
Perplexity / NotebookLM
Real-time web synthesis or deep-dive into your own documents. Two different research postures, both essential.
research analysis sourcing
Video & Motion
Runway / Kling / Sora
Text-to-video, image-to-motion, style transfer. A category that barely existed three years ago, now rivaling production budgets.
video animation production
Workflow & Automation
Make / Zapier / n8n
The connective tissue. Binds your agents together, routes outputs, triggers actions across your entire stack without custom code.
automation integration ops
Features Become Workflows

It started as a
button.
Now it's the whole process.

The first wave of AI in creative software was additive: a "Remove Background" button, a smart crop, an auto-caption. Features you could ignore.

That phase is over. AI isn't a feature in the workflow — it IS the workflow. The tools that survived are those that restructured their entire logic around AI-first thinking, not those that bolted on a smart feature.

The implication for builders: you're not adding AI to your product. You're building a product that wouldn't exist without AI. That's a fundamentally different design problem.

"Why am I asking an agent to do this research for me — when I could build the app myself, own the output, and actually learn something in the process?"

The question that started Trip Atlas
The Case Study

Trip Atlas:
built native to the
AI era.

Trip Atlas is a generative AI web application for photographers and visual travelers — it curates photo spots in any city you're headed to, replacing the hours of pre-trip research most people never have time to do properly.

It replicates the hours a photographer spends before any trip — the searching, the tab-hoarding, the Reddit threads, the note-taking — and compresses it into a curated list of photo spots for any city you're headed to. The research still happens. You just don't have to do it.

AI-first Product Philosophy

Currently curates photo spots by city. The roadmap extends this into full itinerary building — turning a list of locations into a sequenced, shootable plan.

GPT-4o AI Engine

GPT-4o handles the full reasoning load — generating spots, details, coordinates, and context in a single, carefully engineered prompt.

Solo-built Team Size

What once required a funded startup and cross-functional team, Gen AI tools made achievable for a single builder with vision.

AI Engine GPT-4o (OpenAI) generates 10 photo spots per city
Spot Data GPT-4o prompt output name, address, coords, times, gear, food, hours
Geocoding OpenStreetMap Nominatim free country lookup API
Imagery Google Places Photos API stored permanently via Vercel Blob
Origin Claude Sonnet built original React file in Claude Code
UI v0 by Vercel won bake-off vs Lovable
Code assist GitHub Copilot development support
Hosting GitHub version control + PR workflow
Fallback 10 template spots Cathedral, Waterfront, Market, Viewpoint…
The Lab

Eight experiments.
One deliberate
research arc.

Trip Atlas didn't start with a blank canvas. It started with months of structured testing — working through AI tools one by one, understanding their strengths, their limitations, and their behavior under real design and accessibility conditions.

By the time the first line of Trip Atlas code was written, the tool decisions weren't guesses. They were conclusions. All eight experiments ran under the same umbrella: what can Gen AI actually do in a professional design and product workflow — and where does it fall short?

Experiment 1 · UXPilot + Figma
Screenshot to proof of concept
Can AI move directly from a screenshot to a viable proof of concept — skipping layered files and components entirely? A build-first approach using a post-booking lodging screen from Expedia. Key discovery: trying to reproduce the full page introduced noise. Narrowing to one dynamic module — Business Traveler vs Family Traveler — produced cleaner, more usable output. UXPilot didn't just generate UI; it labeled intent. The scope lesson: generate what you need, not what you have.
UXPilotFigmascope
Experiment 2 · v0 by Vercel
Same challenge, different agent
Identical brief to Experiment 1 — but using v0. Introduced the concept of Intent Preview: watching the agent think through the build in real time. v0 rendered live code, not static mockups, enabling real interaction testing (entire card clickable vs. title only). Running out of credits mid-experiment broke momentum and confirmed that iterative AI workflows have operational constraints, not just design ones. Prompt precision matters more in a new environment than in an established one.
v0intent previewlive code
Experiment 3 · v0 + VoiceOver + Silktide
AI accessibility meets real assistive technology
Can AI meaningfully assist in evaluating accessibility when paired with actual screen reader testing? AI surfaced broad WCAG patterns. VoiceOver in Safari revealed how they actually behave — Rotor menus, Tab key navigation, Control+Option arrow navigation. Key insight: not all accessibility issues are structural. Some are environmental (Full Keyboard Access was disabled). Bigger finding: simulated accessibility narration from the AI was idealized — best-case semantics, not real-world implementation. Structural correctness ≠ experiential clarity.
VoiceOverWCAGA11ySilktide
Experiment 4 · v0 + Accessibility Annotations
Annotations as authored intent
Can AI generate accessibility annotations by type — as designed overlays, not post-hoc audits? Color-coded annotation layers generated: heading structure (purple), focus order (blue), screen reader output (green), tab order (black). The agent went further than asked — creating individual URL views per annotation type and a toggle to turn numbers on/off without prompting. Key gap found: screen reader labels drifted from visible UI labels. Technical accuracy ≠ accessibility clarity. An interface can pass checks while still confusing a real user.
annotationsARIAagent initiative
Experiment 5 · Figma Make
Model parity does not equal workflow parity
If multiple agents share the same underlying model family, should the same prompt produce comparable outputs? Tested with Figma AI using the identical brief from Experiments 1 and 2. Result: Figma AI hallucinated a new layout rather than reconstructing the provided screenshot — repeatedly, even after explicit correction. The divergence wasn't subtle. It was structural. Conclusion: orchestration, tool constraints, and product design shape outcomes more than the underlying model. Prompt portability has hard limits.
Figma Makehallucinationtool limits
Experiment 6 · Lovable
Brand fluency vs token substitution
Can an AI agent integrate brand guidelines and operate at a meaningful level of design exploration with minimal prompting? Lovable correctly identified the screen as an Expedia post-booking page — the first agent in the series to do so. Strong structural reconstruction and fast execution. But visual expression was generic. Branding was surface-level: color tokens swapped, brand fluency absent. A11y testing with VoiceOver found unlabeled buttons (WCAG violation) only caught after explicit prompting. v0 leads on visual reasoning. Lovable leads on structural fidelity and context recognition.
LovablebrandA11ycomparative
Experiment 7 · Claude
Design system architecture at scale
Can Claude architect a full, production-ready design system from scratch? Built EDS 2.0 — a comprehensive Expedia Design System with 34 components, 180+ design tokens, WCAG AA+ compliance, full dark mode support, a 9-level type scale, 12-column responsive grid, and complete developer documentation including React usage, CSS custom properties, Style Dictionary token pipeline, and Figma library structure. Claude as a thinking partner for systems-level design — not just a code generator. This experiment established the foundation for everything that followed.
Claudedesign systemtokensWCAG AA+
Experiment 8 · Claude Sonnet
Travel Concierge — ideation at product scale
First substantial product ideation session with Claude — outlining a travel concierge concept that had been in development mentally for several years. Scope: lodging, flight, and car rental across five trip phases, single and multi-traveler scenarios. Claude asked structured clarifying questions before building. Key process innovation: asking Claude to review responses before implementing — a UAT loop that avoided unnecessary churn. Result: Feature Matrix, Task Flows, Disruption Flows, Design Principles. Claude identified one proposed mechanic as "a genuinely novel mechanic — I haven't seen it done well in any travel app." This experiment was the direct conceptual predecessor to Trip Atlas.
Claude Sonnetproduct ideationtask flowsUAT loop

"By the time Trip Atlas started, the bake-off between v0 and Lovable wasn't a guess. It was the conclusion of a deliberate research arc — all under the umbrella of Gen AI."

AI Lab · Experiments 1–8 · jenniferfrench.design/ai-sandbox
The Origin

Not a concept.
A real problem
she needed solved.

Trip Atlas didn't start as an experiment in AI development. It started as a photographer preparing for a trip to several international cities — and running out of patience with the research process.

The hours spent cross-referencing travel blogs, Reddit threads, Google Maps pins, and photography forums — all to answer one question: where should I actually go to shoot? That friction was the brief. The app was the answer.

Trip Atlas was conceived, ideated, and built to provide locations, descriptions, sample photos, geo-related data, access information, equipment recommendations, and nearby places to eat. It was used in the field during actual travel — the ultimate user test.

The Build

From Claude Code
to production
the full journey.

01

"Where to Shoot" — the first prompt

On February 21, the original brief went into Claude: build an app listing the top 5–10 places to photograph in cities entered by the user, with maps, sample images, recommended times, equipment suggestions, and nearby food. Claude named it "Where to Shoot" and built the React file. The four seed cities: Edinburgh, Amsterdam, Barcelona, Dublin.

02

The bake-off: Lovable vs v0

Lovable's concept — named "Trip Muse" — came with a full design vision: "Polarsteps meets Google Travel. Dark theme with warm amber accents, high-contrast typography, subtle grain textures." Visually striking, full-bleed. v0 offered a more maintainable component-based system. v0 won. The amber accent and dark theme survived.

03

GitHub integration

Connecting v0 to GitHub established a professional PR workflow — changes pushed automatically to a feature branch, reviewed, then merged to main. By the end of the build: 146 commits, 106 deployments, 5 active branches. TypeScript 97.9% of the codebase.

04

A/B: tabs vs list

Two information architectures were tested side by side — cities as horizontal tabs with a spot grid below, vs cities as expandable sections in a vertical scroll. Tabs won. Cleaner on mobile, more app-like, and a better fit for the dark photography aesthetic.

"The relative speed of the UI design from v0 — and some of the pattern suggestions it put forward during iteration — genuinely surprised me."

Jennifer French, on building Trip Atlas
Iteration & Testing

Ship, test,
fix. Repeat.

Development moved through multiple revision cycles, each driven by hands-on testing. A structured UAT process was established early — and became one of the most important decisions of the build.

The UAT protocol required the agent to verify changes after each implementation, confirm fixes across all affected screens, and get explicit sign-off before any PR merge. Small discipline, big payoff.

Iteration
Footer addition
Persistent footer with copyright, version badge, and About drawer — added for professional polish and app transparency.
Iteration
Skeleton loader
Replaced a "Location not found" flash with a skeleton loading state. Perceived performance improved significantly without changing actual load time.
Iteration
Privacy messaging
"You will not be spammed" added to auth screens. A single line of copy that meaningfully reduced friction at account creation.
Iteration
Badge repositioning
Archived badge moved from overlapping the 3-dot menu to left-aligned on trip cards — eliminating a UI collision that wasn't visible until testing.
Iteration
Button clarity
"Add Spot" became "Add a custom spot" — a small copy change that eliminated user confusion about what the action would do.
UAT Rule
No merge without sign-off
Explicit confirmation required before any PR merge. Multi-screen changes verified across all affected screens. Rollback procedures defined in advance.
Problem Solving

The images
stopped loading.
All of them.

During beta testing, a critical issue emerged: spot images across every city and every trip stopped rendering after an unknown period. The database looked fine. The URLs looked valid. But Google had quietly expired them.

Google Places Photos API returns temporary URL references — not permanent image files. Once expired, every stored URL became a broken link. A trip created in March 2026 showed broken images weeks later across all 40 spots in Edinburgh, Amsterdam, Barcelona, and Dublin.

Option considered
Refresh on view
  • Fetch a new URL each time a spot is viewed
  • Always current, never broken
  • Ongoing API costs at $7 per 1,000 requests
  • Added latency on every page load
Solution chosen
Vercel Blob storage
  • Download and store actual image files permanently
  • One-time implementation cost
  • Images never expire — ever
  • ~$0.02–0.25/month ongoing storage cost

"I had to keep purchasing credits mid-debug. Eventually I instructed the agent to roll back and phase the builds in — that was the real turning point."

Jennifer French — the lesson that changed how she worked with the agent
Honest Reflection

Building fast means
the process
disappears.

One of the unexpected costs of building at AI speed: you're moving too fast to document what you're thinking. The case study came after the build — which means a lot of the in-the-moment reasoning had to be reconstructed from chat logs and commit messages.

Now multiply that across several agents, each handling a separate task — one writing code, one generating UI, one managing version control, one sourcing images. Each conversation its own context, its own decisions, its own dead ends. The process doesn't just disappear once. It disappears in parallel.

This isn't a failure of process. It's a feature of the medium. When you're iterating at the pace Gen AI makes possible, reflection is the thing that falls away the easiest. The work moves faster than the meta-awareness of the work.

The lesson for anyone building in this era: decide early whether you're making a product or a case study. If it's both, you need to slow down enough to capture the thinking — because the tools won't do it for you. At v115, mid-debug, with credits running out, nobody is writing field notes.

150+ GitHub commits

Each one a decision. Most of them made too fast to remember why.

100+ Builds & deployments

Ship, test, break, fix, repeat. The rhythm of AI-assisted development.

v115 When credits ran out

Mid-debug. The agent stopped. Credits purchased. Work continued. No documentation written.

Lessons Learned

What building
in the AI era actually teaches you.

01

Make the agent plan before it builds

The single biggest process improvement: before any implementation, require the agent to think through the build and generate a plan. What it's going to do, in what order, and why. This one step eliminates entire categories of downstream bugs — and gives you something to review before a line of code is written.

02

UAT after every build, before every push

Instruct the agent to conduct UAT immediately after implementation — before the PR is created, before anything merges. Bugs caught at the build level cost tokens. Bugs caught after deployment cost time, trust, and sometimes a full rollback.

03

It's not the scope. It's the bug you didn't catch.

Somewhere inside a large build, something broke quietly. Not dramatically — just a small, undetected fault that subsequent work stacked on top of. By the time it surfaced, it was load-bearing. The scope didn't cause the problem; it just gave the problem more places to hide. Building with AI agents is a series of levers and pulleys — each action triggers the next, and the learning happens in the chain reaction. The fix isn't smaller ambition. It's earlier detection. Which is why UAT isn't optional.

"At the end of the day it's a Rube Goldberg machine — levers and pulleys, give and take, and the learning happens somewhere in the chain reaction."

Jennifer French
04

Your taste, your empathy, your humanity — these are what the agent can't replicate.

When generation is cheap, curation is expensive. But it goes deeper than taste. It's the empathy to ask: how does this actually feel when a real person uses it? Technically everything works in theory. The skeleton loader loads. The UAT passes. The PR merges clean. And then someone uses it in the field — tired, distracted, on a phone in bad light — and you realize the gap between "works" and "works for a human" is where all the real design lives. That gap is yours to close. No agent gets there without you.

What's Next

The methodology
is compounding.

Trip Atlas isn't the destination. It's where the approach was proven. What comes next is the same method, applied with more confidence and fewer wrong turns.

Trip Atlas · Roadmap
Itinerary building
The spots already know their hours, access details, and ideal shooting windows — golden hour, blue hour, when the crowds thin. The itinerary builder takes it further: starting from where the user is, understanding their preferred first shot of the day, then sequencing the full day around their mode of transport. Location-aware, light-aware, human-aware.
Next Build · In Progress
Lightroom Classic
A workflow plugin for Lightroom Classic — currently in early development. Same AI-assisted build methodology as Trip Atlas, with low-fi wireframing from the start. More details to follow.
Evolving Method
Low-fi first
Already finding efficiencies: low-fidelity wireframing in early stages instead of jumping to hi-fi. Less time spent on polish before the structure is right. A small shift that changes everything downstream — and one that only comes from having done it the hard way first.
Final Thought

We are not at the end
of creativity. We are at
its reinvention.

Trip Atlas is one answer to the question Gen AI is posing to every creative industry: not "will AI replace you?" but "what will you build now that you can?"

© 2026 Jennifer French. Licensed under CC BY-NC-ND 4.0. You may share this work with attribution for non-commercial purposes. No derivatives permitted.