Key Takeaways
- AI coding agents have matured to the point where experienced designers can build production applications without framework-specific coding skills.
- The best development workflow focuses on data models, transaction design, and user experience — then lets AI handle implementation.
- Small businesses benefit most — custom, high-quality applications are now accessible without enterprise development budgets.
- Design expertise is the real bottleneck, not the ability to write code in a specific framework.
One App, Three Frameworks, One Saturday
I spent today rebuilding the same application three times.
It's a car maintenance logger — a straightforward business tool that tracks service history across multiple vehicles. Nothing glamorous, but it's the kind of practical utility app that small businesses need and rarely get built well.
The first version was React with Ant Design. Functional, solid, the kind of application I've been building for clients for years. Then I asked my AI coding agent — GitHub Copilot running Claude — to rebuild it from scratch in Vue 3 with shadcn-vue and Tailwind CSS. Same database, same PHP API, completely new frontend. It was built, debugged, and deployed to production in a single session.
Then I started asking harder questions. Not "can AI build this?" — that's been answered. But "what is the absolute best technology stack for this type of application, given my deployment constraints?" That question led me down a fascinating rabbit hole of framework research, and ultimately to a third architecture: Svelte 5.
Here's the part that matters: I didn't write a single line of framework code. I designed. I specified. I reviewed. I directed. The AI wrote every line of Vue, every TypeScript interface, every Tailwind utility class.
The Problem I've Been Solving Wrong
I've been designing user interfaces since 1992. My first professional UI work was on IBM 3270 mainframe terminals — green-on-black text screens where every keystroke mattered. From there I moved through seventeen years of designing interfaces for blue-chip clients worldwide as a usability architect, then founded Dygiphy in 2009 to serve Australian small businesses.
In those enterprise roles, the workflow was clear. I'd research users, design the interaction model, specify the transaction flows, and hand detailed specifications to a development team who knew their tech stack inside out. Java developers built Java. .NET developers built .NET. My job was to make sure the end result was excellent for the user, regardless of what was happening under the hood.
When I went solo, I lost that luxury. Suddenly I wasn't just the designer — I was also the developer, the project manager, the infrastructure engineer. And as a developer, I was constrained by the frameworks I personally knew well enough to be productive in.
That's backwards. The technology should serve the design, not constrain it.
Why I Rebuilt the Same App Three Times
This wasn't wheel-spinning. It was methodical evaluation.
The React build used Ant Design — a comprehensive component library that gives you everything out of the box. Date pickers, form validation, data tables, charts. The developer productivity argument is strong: you assemble pre-built components rather than building from scratch.
But that productivity argument has always had a cost. Ant Design ships a significant JavaScript payload. The React runtime adds more. The total bundle for a relatively simple application sits well north of 350KB gzipped — and my clients' customers are often on mobile connections in regional Australia.
The Vue build with shadcn-vue and Tailwind CSS came in at roughly 230KB total. Same functionality, substantially smaller footprint, far more design control. shadcn-vue gives you unstyled Radix primitives that you compose yourself — the polar opposite of Ant Design's opinionated component library.
That result made me curious. If moving from a heavy library to unstyled primitives cut the bundle by a third, what would happen with a framework that compiles away entirely?
Researching the Right Stack
I dug into the data. The 2024 State of JavaScript survey shows Svelte topping the satisfaction and retention rankings — developers who use it overwhelmingly want to keep using it. The Stack Overflow 2024 Developer Survey reports 73% of Svelte developers want to continue using it, and Vite — the build tool that powers all three of my implementations — has the highest admiration score of any build tool at 77%.
Svelte 5 compiles components to vanilla JavaScript at build time. No virtual DOM. No framework runtime in the browser. For applications deployed to shared LAMP hosting — which is the reality for most of my small business projects — this delivers the best possible performance with the smallest possible payload.
My final stack recommendation: Svelte 5 with runes, shadcn-svelte for UI primitives, Tailwind CSS 4 for styling, and Motion for animation. All compiled to static assets that deploy anywhere a web server can serve HTML, CSS, and JavaScript.
Transaction Design — What Actually Matters
Here's where it gets interesting.
Once I'd settled on the stack, I wrote a comprehensive specification for the Svelte rebuild. Not code — a specification. And I made a deliberate choice about what to include and what to leave out.
What I specified:
- Complete database schema — six tables with every column, constraint, and relationship
- Full API contract — every endpoint, every request format, every response shape
- Authentication flow — cookie-based persistent tokens, middleware behaviour
- Environment configuration — how to detect local vs. production, database credentials
- User workflow descriptions — what the user is trying to accomplish and in what sequence
- A UX design brief — focused on workflow efficiency, data density, and touch targets
What I deliberately left out:
- Component names
- Page layouts
- Navigation patterns
- Colour palettes
- Typography choices
I've always called this approach "transaction design" — defining the data model and the transactions the user needs to perform against it. It's what I was trained to do during my Computer Science degree in 1992, and it's what I've done for every project since. The visual design is important, but it flows from the transaction model, not the other way around.
By giving the AI complete creative freedom over the visual and interaction design — but iron-clad constraints on the data model and business logic — I'm leveraging what each party does best. I understand the user's workflow intimately. The AI can synthesise current best practices in component composition and visual design faster than I can manually research them.
The Shift That Made This Possible
The Stack Overflow 2024 survey reports that 76% of developers are using or planning to use AI tools in their development process — up from 70% the previous year. Among those using AI tools, 82% are using them to write code. The top cited benefit? Increased productivity, at 81%.
GitHub's developer experience research found that 92% of developers are already using AI coding tools at work. Critically, developers believe these tools give them more time to focus on solution design — exactly the shift I experienced today.
Jakob Nielsen — the researcher whose work has informed my design practice for three decades — describes this as the first new UI paradigm in 60 years. He calls it a shift from command-based interaction ("do what I say, step by step") to intent-based outcome specification ("here's what I want — figure out how"). It's the difference between writing code line by line and describing what you want the application to achieve.
For someone with my background, that framing resonates deeply. I've spent my entire career specifying intent and outcomes — that's literally what user-centred design is. The tools have finally caught up with the methodology.
What This Means for Small Business
Custom business applications have traditionally been expensive. A bespoke tool — even a simple one — might cost $20,000 to $50,000 and take months to deliver. You're paying for developers who need to understand your domain, learn your data model, make architectural decisions, and then implement them in whatever framework the agency happens to specialise in.
That cost has pushed most small businesses toward off-the-shelf SaaS products that almost-but-don't-quite fit their workflow. They adapt their processes to match the software, rather than the other way around. Every workaround is a friction point. Every missing feature is a daily frustration.
AI coding changes this equation fundamentally. The expensive part of building custom applications was always the implementation — finding developers with the right mix of framework expertise, domain knowledge, and design sensibility. That combination was rare and getting rarer.
With AI handling implementation, the bottleneck shifts to design expertise — understanding what the user actually needs, structuring the data model to support those needs, and specifying the transactions clearly. That's a skill set I've been building since before the web existed.
My clients don't need a React developer or a Svelte developer. They need someone who understands their business workflow and can translate it into a well-designed application. The framework choice becomes a technical detail that serves the design, not a constraint that limits it.
Thirty Years of Design Experience, Unshackled
This is genuinely the most exciting moment in my career. Not because the technology is impressive — although it is. It's because for the first time since I left enterprise consulting, I can design applications the way I was trained to.
Start with the user's workflow. Work back to the data model. Specify the transactions. Then hand the implementation to something that can execute it faithfully — in whatever framework best serves the end user.
I'm no longer constrained by which frameworks I personally know. I'm no longer compromising designs because I can't find or afford a developer with the right tech stack experience. I can choose the best possible technology for each job and trust that the AI can implement it competently.
The result? My small business clients get custom applications that fit their exact workflow, built with the best available technology, delivered in days instead of months. Applications that make their everyday work easier and more rewarding to do.
That's always been the goal. The technology was never the point — it was always the means. AI has finally made it possible to treat it that way.
