The problem
Freelance web design usually forces a tradeoff: move fast and risk a generic, fragile product, or move slow and price yourself out of most budgets. I wanted a third option. I wanted a system flexible enough to scale timeline and cost to what a client actually needs, without sacrificing structure, accessibility, or the kind of UX thinking that makes a site easy to live in long after launch.
The approach
I use Claude as a true design/development partner…not a vending machine for finished pages.
It starts with the concept: the goal, plus any research or discovery data fed to Claude Code to generate a first-pass site. It’s usually quick and good-looking, but I never treat it as the deliverable. If I don’t understand the foundation it’s built on, then I can’t trust it or use it long-term, and neither can my client.
So I screenshot the output and bring it into Figma, where I rebuild it as wireframes, from low fidelity up to high fidelity. I use the AI draft as inspiration, not gospel. I audit each version, then start constructing the actual design system and components from that audit.
Here’s a live website that AI generated for me in seconds:
And here’s where I took that and ran with lofi’s:
And the UI Audit helps me spit out a Figma-based design system file where I can begin to create the branding, component library, and high-fidelity pages:
From here, I train Claude Code on my design system and token usage, and build component by component, feeding it screenshots and specific interaction requirements rather than vague prompts. Once I’m in the codebase, I work directly in the terminal: sometimes letting Claude Code make edits automatically, sometimes turning that off entirely so I can review line by line.
I use my Figma component properties panel to guide the AI in the creation of the component’s code, and tweak the code when it inevitably makes a mess of things.
That last part matters most. I’m not a developer by trade, so the second I lose track of what’s doing what in the code, the whole foundation starts to feel shaky. Staying hands-on with every layer, from the design system to components to code, is what keeps a fast process from becoming a fragile one.
The outcome
The result is a workflow where AI compresses the timeline dramatically, but a human (me, the real design expert) is still the one making every structural and design decision. That’s what lets me offer real tiers (1 day, 1 week, 1 month) based on client needs, without ever shipping “some whatever website.” Every site that comes out the other side is solidly structured, fully editable by non-technical clients without breaking, and built with accessibility and UX in mind from the start. Not to mention, confident support from their trusted creative problem solver (again, me).
The bigger takeaway
AI didn’t replace the design process here. It just changed the pace. The interesting part isn’t that a model can generate a site in seconds; it’s figuring out where human judgment needs to stay in the loop so speed doesn’t come at the cost of quality. That’s the actual skill I’m building with JB & Co.: not prompting, but knowing exactly where to insert myself in an AI-accelerated world.
