My actual AI workflow after 22 years in design
Francois Brill
Founding Designer
I've been building digital products since 2003. Several businesses. Countless client projects & products. Every tool transition the industry has produced.
And now, AI.
Here's the thing: I'm not going to tell you AI changes everything. I'm also not going to tell you it's overhyped. Both takes are lazy.
What I'll tell you is exactly how I use it - the specific tools, the specific workflows, the specific places where it helps and where I still do things by hand.
The tools
Let me just lay out the stack:
Figma — Still where design thinking happens. Design systems, components, layouts. But now set up specifically to mirror Tailwind config, because of what comes next.
Figma MCP + Cursor — The bridge from design to code. Plug Figma's MCP server into Cursor, and I'm getting 80-90% accurate code from designs on the first pass.
Cursor — Daily driver for coding. Composer mode for normal tasks, tab completion for flow state. Custom rules to map AI outputs to our actual Tailwind variables.
Claude Code — Deeper thinking. Plan mode for complex problems. Runs locally, which means I can have it iterate on content inside the actual files.
Granola — Transcribes every client call. Not for playback—for pattern recognition across hundreds of conversations.
That's the most important parts of the stack. At least for now, late 2025 this is how the job get's done, most certainly this will change and evolve, so don't take it word-for-word, take it for what it is in the current toolset and what the pieces fit together.
Now here's how it actually works.
Speed vs. unlock
AI does two different things for me, and it's worth separating them:
Speed: Same work, dramatically faster. Figma to code in hours instead of days. Component variations in minutes instead of hours. Iteration cycles compressed.
Unlock: New work that wasn't practical before. Capabilities I couldn't offer, dependencies I couldn't remove.
Most people focus on speed. The unlocks are more interesting.
“Most people focus on speed. The unlocks are more interesting.
Unlock #1: Copywriting without the bottleneck
Traditionally, copy was a dependency. I'd build a landing page with placeholder text, then wait. For the client. For their copywriter. For revisions. The design would sit there, incomplete, hard to evaluate.
Now I generate solid first-draft copy as part of the design process. Product pages. Feature descriptions. UX microcopy. Even if it gets revised later, the design is complete—stakeholders can react to something real.
This isn't replacing copywriters. It's removing a blocker that used to stall projects for weeks, and for some clients this is better copywriting than they had in the first place.
Unlock #2: Learning from your own history
We've been using Granola for over a year now. Every discovery call, every client meeting, transcribed and searchable.
Here's what that enables: I can search horizontally across all our conversations. What pain points keep coming up? What language do clients actually use when talking about our value or their pain points? Where do we deliver the most value in their minds?
This is research I couldn't practically do before. The data was always there - in my memory, fragmentary and biased. Now it's searchable.
The Figma-to-code reality
Here's the speed improvement that's actually changed our delivery time:
Set up Figma properly—design system mirroring Tailwind, auto layout everywhere, variables for colors and spacing. Sweat some details upfront.
Connect Figma MCP to Cursor. Build the screen in minutes.
First pass: 80-90% accurate. Components in place, layout correct, styles applied.
The gap? AI sometimes hardcodes values instead of using variables. I've written specific Cursor rules to automatically map those back to our Tailwind config. Another layer automated.
One or two rounds, and we're pixel-perfect.
What's left for humans: image optimization, responsive refinements, the judgment calls. But the translation from design to code—the part that used to take days—now takes hours.
Why I still write code by hand
Here's what might surprise you: I still manually write a lot of code.
Styling. Layout adjustments. Getting things positioned exactly right. Responsive behavior. These are faster for me to do directly than to describe in prompts.
But Cursor's tab completion changes everything. It's autocomplete at the speed of thought. I type, it predicts, I tab to accept. My decisions, my code, just a lot faster.
This is flow state, augmented. Not automated—augmented.
For the work that requires taste, for the details that need human judgment, I want my hands on the keyboard. AI keeps pace so I don't have to slow down.
“This is flow state, augmented. Not automated—augmented.
Plan mode is where thinking happens
Here's my most important AI workflow: using plan mode before building anything complex.
Claude Code and Cursor both have versions of this. You describe what you want to build. The AI asks clarifying questions. You think through the problem together. Then you lock in a plan and execute.
This isn't outsourcing thinking—it's structured thinking with a collaborator that can execute instantly once you've decided.
The key is staying in the loop. Let AI ask questions. Make the decisions yourself. Then hand off execution.
You're still the one deciding. AI just runs faster than typing.
It's also worth mentioning that I use voice dictation on a lot of these thinking sessions, as it's a lot faster than typing and getting a brain dump out of your head and into the AI context. It's all about providing the right context.
Content in the codebase
Here's a workflow that's been transformative: generating content as part of the build, not separate from it.
Because Claude Code runs locally, I can have it iterate on copy inside the actual files. Build a perfect feature page template. Then tell Claude to generate the other pages—structure, layout, and copywriting included.
It understands the pattern. It produces pages coherent with the system.
This is "human designs the system, AI extends it." The craft is in the system design. The execution scales.
“Tools change constantly. The judgment required to use them well doesn't get automated.
The judgment layer
After 22 years, here's what I know:
Tools change constantly. The judgment required to use them well doesn't get automated.
AI makes me faster. It removes blockers. It handles mechanical translation.
But every meaningful decision is still mine. Which problem to solve. Which direction to pursue. When something is right versus almost-right. How to build systems that cohere.
I use AI all day, every day. And I've never been more certain that craft—human judgment, applied with intention—is what actually matters.
The tools are better than ever. The work is the same: good decisions, made well.
