How I design with AI
I used to hand my engineers a static figma frame and hope.
Hope the hover state was obvious. Hope the empty state made sense. Hope the thing in my head survived the trip into code.
It mostly didn't. A flat frame can't carry interaction, and everything interesting in a product lives in the in-between. That gap is why I started designing code alongside the screens. I still work in Figma, but now I can design the interaction and the implementation together.
Agent-native, not augmented
AI is the pink elephant in every design conversation right now, and you're expected to have adopted it. Mine never came from pressure. It came from one question: could this actually make the work better? Speed up the slow parts. Smooth the research. Leave more room for thought and craft.
There's a name for where it landed: agent-native, not AI-augmented.
The difference is the whole thing. AI-augmented is a tool you reach for inside the old loop. Agent-native is the loop changing shape. The agent isn't something I use, it's something I brief. I set direction, it produces, I review, I decide what ships.
My attention moved up a level. From pushing pixels to judging whether what came back is good enough.
Abstraction is cheap, so here's the actual week.
Vibe coding the portfolio
The portfolio you're reading was vibe coded. Next.js, Tailwind, built in codex and claude code. I designed the homepage and case study template in figma, then built it out with the agent, component by component.
I'm a designer, not an engineer. A year ago that sentence ends the conversation. Now it just means I review pull requests instead of writing the code from scratch. I still make every call that matters: layout, hierarchy, motion, the kerning on the heading you just scrolled past. The agent handles the part that used to need a front-end developer and a two-week queue.
The team playground
I built a playground my whole team can step into. Anyone can open a working version of our product and click through it. It earns its keep the moment static designs need to become code: my engineers see how something actually feels, and we surface the edge cases a flat frame quietly hides.
Handoff stops being a document I throw over the wall. It becomes something we can both poke at.
Colours without hex
My colours never touch hex in code. Hex is an address, not a description. #4A90D9 tells you nothing about what it is or what sits next to it, so every palette becomes a set of values you hand-picked once and can't reason about again.
OKLCH describes the colour instead: lightness, chroma, hue. Three numbers that mean something. Hold lightness and chroma steady, spin the hue, and every colour in the set carries the same visual weight. Step the lightness down and you get a shade ramp that stays the same colour the whole way, no purple creeping into the tints, no muddy grey in the darks. Perceptual uniformity, which is a dry phrase for the thing designers have always eyeballed by hand.
So the pipeline runs from figma straight through to the build, and the design file is the source of truth. No drift. No "designer says one thing, code says another". Change a token in figma and it changes everywhere, and the agent does the wiring.
This is the unglamorous infrastructure AI is genuinely brilliant at. What used to eat an afternoon of copy-pasting now happens while I think about something more interesting.
Skills as taste
I built my own skills for the boring-but-crucial passes. An animate skill. A delight skill. A polish skill. Small custom workflows I dispatch when a page is structurally done but not yet alive.
They encode taste I've already articulated once, so I don't re-explain "use xs-shadow on the secondary button" every single time. This is the move I care about most. When a design decision is portable, it becomes a rule the agent applies, not a thing I redo by hand.
The second brain
The whole system sits on a second brain. I run a vault that claude code treats as a chief of staff. My notes, my goals, my voice, my project context, all readable. So when I dispatch work, the agent already knows how I write, what I'm building toward, and which decisions are settled.
The context isn't stuck in my head where it bottlenecks. It's on disk where the agent can use it.
Taste is the scarce thing
Every hiring conversation right now is circling the same fear: that this replaces the craft.
It does the opposite. When production gets cheap, taste becomes the scarce thing. The agent will happily build whatever I brief, including something mediocre, and it will build it fast, confidently, and with good commit messages. The quality of my judgement is now the entire game.
That's the part that never got automated.
I'm deliberate about what I hand over. Early exploration, the messy thinking where I don't yet know what I want, stays mine. AI is bad at taste and brilliant at execution, and the whole skill now is knowing which is which. The rest I dispatch fully, and it's bought back the one thing designers never have enough of: time on the decisions that actually move the work.
The bar didn't drop. It rose. AI just took the floor out from under the people competing on output, and put the whole weight on craft.
