Thinking

The judgment behind AI builds

This site was built in conversation with AI, plus code I refined. These are the decisions I made along the way, with the commits that prove them.

Stack
Claude CodeCursorVS CodeConductorOmnigentChatGPTPerplexityGitHubVercel

3D+1,770 −16

Creative exploration

The professional site stays intact. A second skin shows range: holographic, playful, hand-built.

Hover to tilt · click to turn

components/fun/HoloCards.tsx
// A DOM enhancer, not a wrapper: images live in many different containers// (next/image fills, lightbox buttons, screen-flow buttons, plain clipped divs).// Find each image's visual card, tag it, write transform directly. React never// sets style on these nodes, so the two don't fight. Undone on unmount.const SEL = "main img, main video" function scan() { const main = document.querySelector("main") const seen = new Set<HTMLElement>() main.querySelectorAll<HTMLElement>(SEL).forEach(media => { const card = findCard(media, main) // nearest clipping ancestor if (card) seen.add(card) }) seen.forEach(attach) for (const [el, c] of cards) if (!seen.has(el) || !el.isConnected) { c.off(); cards.delete(el) }} // Gesture model: TURN owns the card (left-drag, momentum, settles to a face).// MOVE is secondary: grip bar, press-and-hold ~0.4s, or shift / right drag.// CLICK collects. No hint; the tray appears after the first collect.
Asked
Give the site a second skin: holographic, three-dimensional, a little game-like, without changing a word of content or adding a WebGL scene. Every image should become a card you can turn and move. Make discovering it part of the fun.
Came back
The first pass wrapped each image in a new card component, which meant touching every container on the site and keeping two trees in sync. I asked for a DOM enhancer instead: on entering the mode it finds each image's visual card, tags it, and writes transforms directly; React never styles those nodes, so the two never fight, and everything is undone on exit. The gesture model was the other call. Turn is the primary action and owns the whole card, the 3D-viewer convention; move is secondary and never competes for the same gesture (a grip bar, press-and-hold, or shift-drag); a click collects. Nothing announces the game. The tray only appears after the first collect.
Kept
One switch reboots the site into a dark holographic world: a canvas starfield, every image a card with a wireframe back face carrying its caption, momentum on the turn, a collection tray, and levels that recolour the sky. Plain canvas and CSS 3D. Zero new dependencies.

Navigation+40 −27

Focused, but curious

A mode switch should make people curious, not compete with the nav.

Hover it · click to switch this box

The orbit turns on its own. The label only appears when you ask, in the same mono caps as every eyebrow on the site.

components/FunToggle.tsx
<Orbit className="fun-toggle-orbit h-4 w-4 lg:h-[1.125rem] lg:w-[1.125rem]" /> {/* Typed in: each glyph fades up in sequence, in place, so nothing slides. */}<span aria-hidden className="fun-toggle-label absolute right-0 top-full mt-1.5 font-mono text-[11px] uppercase tracking-[0.18em]"> {label.split("").map((ch, i) => ( <span key={i} className="inline-block opacity-0 translate-y-[3px] transition-[opacity,transform] duration-150 ease-out group-hover:opacity-100 group-hover:translate-y-0 group-focus-visible:opacity-100 group-focus-visible:translate-y-0" style={{ transitionDelay: `${i * 28}ms` }} > {ch === " " ? "\u00A0" : ch} </span> ))}</span>
Asked
The 'Open portal' pill in the nav is out of step with the rest of the site. Find something quieter that still makes people curious, and keep it consistent with the nav's type.
Came back
The options came back as a restyled pill, a plain text link, and an icon. I took the icon alone: an orbit that turns slowly on its own, so it reads as alive without asking for attention. The label was my addition. On hover or focus it types itself in beneath the icon, one glyph at a time, in the site's small mono caps, the same voice as every eyebrow on the site, so it reads as a readout rather than a tooltip. It is absolutely positioned so the nav row never shifts by a pixel. The copy is a round trip: 'Open portal' going in, 'Close portal' coming back, and the reboot transition between the two literally is one.
Kept
Icon-only orbit in the desktop nav, label on hover and focus, 28ms per glyph. The pressed state and an explicit aria-label carry the meaning for assistive tech, and the spin stops under reduced motion.

Layout+173 −80

Reading without gaps

An empty column reads as broken, so the layout tightens as if the image was never there.

Click the box

app/globals.css
/* The row is a grid of two tracks; a collected card hands its track back. */.project-row { display: grid; grid-template-columns: minmax(0, 7fr) minmax(0, 5fr); column-gap: 4rem; align-items: start; transition: grid-template-columns 0.9s var(--reflow-ease), column-gap 0.9s var(--reflow-ease);} html[data-mode="fun"] .project-row:has(.holo-card.is-collected) { grid-template-columns: minmax(0, 12fr) minmax(0, 0fr); column-gap: 0;} /* The card closes its own box so the row tightens vertically too. */html[data-mode="fun"] .holo-card { transition: height 0.9s var(--reflow-ease); }html[data-mode="fun"] .holo-card.is-collected { overflow: clip; min-height: 0; } /* Siblings settle through a short blur: one movement, not a snap. */.is-reflowing > :not(.project-media) { animation: holo-reflow 0.95s cubic-bezier(0.4, 0, 0.2, 1) both;}@keyframes holo-reflow { 0% { opacity: 1; filter: blur(0); } 32% { opacity: 0.45; filter: blur(3px); } 100% { opacity: 1; filter: blur(0); }}
Asked
After a card is collected, the section should reflow as if the image was never there: the copy takes the full width, on the homepage rows, the hero, and the case-study blocks. Smooth, not a jump.
Came back
The first pass hid the image and left its grid track standing, so the text sat beside an empty column. I had it animate the grid itself. The media column interpolates from 5fr to 0fr and the row gap to zero, the card closes its own box over 0.9 seconds on the same ease the rest of the site uses, and the surrounding copy settles through a short blur so the whole thing reads as one movement rather than a layout snap. The same choreography covers the hero portrait and the annotated case-study blocks, which slide their lists left to fill the space. Every image on the homepage was also brought to one size, so the rows read as a set before any of them move.
Kept
CSS grid tracks that transition to a true zero, a card that closes its own height, and a reflow animation on the siblings. Release brings everything back the same way. Reduced motion skips the choreography and just reflows.

Interaction+45 −8

Feedforward first

Feedback says what happened. Feedforward says what is coming: the reward card wears the new colour before the sky does.

Collect enough cards

Level 1 · Observer

components/fun/collection.ts
/** The level the page should LOOK like. While a reward card is pending the * page stays on the previous level; the card previews the next; the world * catches up on dismiss. */export function shownLevel(st = state) { return st.pendingLevel ? st.pendingLevel - 1 : st.level} function applyPerks() { document.documentElement.dataset.perks = perksFor(shownLevel()).join(" ")} dismissReward() { state = { ...state, pendingLevel: 0 } applyPerks() // the colour shift happens here, not at unlock emit()}
Asked
Collecting cards should unlock levels, each with a visible perk. Levels should feel earned and not countable, and the second one should not be more of the first.
Came back
The first pass flipped the accent colour the instant a threshold was crossed and only then showed the reward card, so the world changed before the player was told why. I reordered it. The page holds the previous level's colours while the card is up; the card itself is drawn in the incoming accent, so it previews the shift it announces; the world follows on dismiss. The tiers were also split so each is its own event: a violet sky, then a magenta sky with a cursor trail, then every card turning gold. The first level comes fast, at two cards; after that the gap is random, three to five, so the next surprise cannot be counted down to.
Kept
Three tiers: Observer, Curator, Archivist. Reward first, colour second. Releasing the cards keeps the level; Reset clears it. The whole state persists per browser, so the hunt continues across pages and visits.

Conversational AIapp/api/chat/route.ts+1 more

Prioritize the question

Efficiency and clarity: answer what was asked, phone-length, then one obvious next step.

Is she actually as good as she claims?

Her CPTO at Remitly publicly recognized her as the lead designer on the ChatGPT launch, which is about as heavy as an endorsement gets. Her direct manager's recommendation says the same thing in more detail. The Remitly case study is the place to see the work itself.

Remitly · ChatGPT case study
app/api/chat/route.ts
Voice- Speak with conviction, not attribution. Never say "she says X". A best friend doesn't quote you. They just know you and say the thing.- Open every answer with a complete, conversational sentence. Never a clipped fragment that could read as curt.- Keep answers concise, usually 2 to 4 sentences. Lead with the strongest point. Hard rules about links- NEVER print raw URLs, file paths, slugs, or routes. The portfolio renders clickable buttons below every answer. Trust those buttons. YES: "Start with the Remitly case study to see the AI work." NO: "Start with the Remitly case study at /work/remitly." // The UI side of that contract: chips are keyed off names, not paths.const FOLLOWUP_MAP = [ { pattern: /\bremitly\b/i, label: "Remitly · ChatGPT case study", href: "/work/remitly" }, { pattern: /\bresumes?\b/i, label: "Download resume", href: "/Ann_Lei_Resume_2026.pdf" }, { pattern: /\b(email|reach out|get in touch)\b/i, label: "Email Ann", href: "mailto:…" },]
Asked
Build a chatbot for recruiters and hiring managers. It should answer like someone who knows me well, keep answers phone-length, never dump links, and move a serious reader toward email or the resume.
Came back
The first system prompt was a résumé with a persona on top: 'she says' in every other sentence, raw URLs in the answers, paragraphs that would not fit a phone. I rewrote it as voice rules rather than a character. Conviction, not attribution; a full sentence to open, never a clipped fragment that could read as curt; two to four sentences; the answer first, then the pointer. The bigger design call was taking navigation out of the model entirely: it names a destination in prose, 'the Remitly case study', 'her resume', and the UI renders the chip, so a hallucinated path can never ship. Colleagues' private kudos may be cited as patterns, never by name. Compensation, off-topic, and unknowns each have a fixed answer.
Kept
Streaming replies from Claude Haiku 4.5 with a cached system prompt, follow-up chips derived from what the answer mentions, and three unasked question chips when there is nothing to navigate to. It has one job: make a busy reader's next step obvious.

Interactioncomponents/Hero.tsx+1 more

A little intrigue

A button that leans toward the pointer invites a second look. One that chases it reads as a trick.

Hover the button

Magnetic pull at 35% of cursor offset · global cursor expands via data-cursor-morph

components/Hero.tsx
function MagneticButton({ href, children, className }) { const x = useMotionValue(0) const y = useMotionValue(0) const sx = useSpring(x, { stiffness: 200, damping: 12, mass: 0.1 }) const sy = useSpring(y, { stiffness: 200, damping: 12, mass: 0.1 }) function onMove(e: React.MouseEvent<HTMLDivElement>) { const r = e.currentTarget.getBoundingClientRect() x.set((e.clientX - (r.left + r.width / 2)) * 0.35) y.set((e.clientY - (r.top + r.height / 2)) * 0.35) } function onLeave() { x.set(0); y.set(0) } return ( <motion.div style={{ x: sx, y: sy, display: "inline-flex" }} onMouseMove={onMove} onMouseLeave={onLeave}> <a href={href} data-cursor-morph className={`inline-flex items-center pl-[2px] pr-8 py-3 rounded-full ${className ?? ""}`}> {children} </a> </motion.div> )}
Asked
Build a magnetic 'View case studies' button that feels tactile and precise. The interaction should support the intent to explore, not perform craft. Keep it readable, accessible, and cheap to render.
Came back
The first pass pulled at half the cursor delta and the button felt like it was chasing the pointer: the label drifted off its visual anchor and the focus ring wandered with it. I cut the pull to 0.35. Lower and the response is invisible; higher and it reads as a trick. At 0.35 the button stays inside its own footprint, so the label, the hit target, and the focus ring all stay put. The spring was retuned for precise on entry and calm on exit, and the global cursor was given a morph attribute so it expands to the button's shape at the same moment the button leans in. They meet partway.
Kept
Two motion values through one spring, transform-only so it stays smooth on weak hardware, and a cursor that wraps the button as the button pulls toward it. Responsiveness reads as a property of the interaction, not a performance of it.

Layoutcomponents/Projects.tsx+1 more

Strongest work first

A 50/50 second row reads as a tie. A 7/5 split keeps a primary and a secondary.

Compare the layouts

Remitly
UoP
MSCI
MS

Strongest project goes full-width to lead reading order: secondary projects compress to a 7/5 row, container widens to feel more editorial.

components/Projects.tsx
@@ grid layout @@- const colClass = [- "col-span-12 sm:col-span-7",- "col-span-12 sm:col-span-5",- "col-span-12 sm:col-span-5",- "col-span-12 sm:col-span-7",- ][i]+ const colClass = [+ "col-span-12",+ "col-span-12 sm:col-span-7",+ "col-span-12 sm:col-span-5",+ ][i] @@ section container @@- <div className="max-w-3xl mx-auto px-6">+ <div className="max-w-5xl mx-auto px-6">
Asked
Rework the work section so the strongest project gets full-width placement and the secondary projects sit in a denser layout below. It should feel editorial without losing density.
Came back
The first proposal put the strongest piece on top and split the second row 50/50. I pushed back: hierarchy has to hold inside a row, not just between rows, or the second row reads as a tie. We settled on 7/5. The other change was mine to call: the content shell was too narrow for editorial pacing to exist at all, so it widened from a reading column to a magazine measure, and the spacious quality came from the negative space around each piece rather than from the pieces themselves.
Kept
Reading order is set before the visitor scans: Remitly as one unbroken composition, University of Phoenix and MSCI in deliberate asymmetry below, inside a wider shell every section now shares.

Accessibilitycomponents/Projects.tsx

Legible over any image

Text over photos is tuned against the brightest, busiest image, with margin over the AA floor.

Toggle the WCAG pass

RemitlyConversational AI

Money Transfer on ChatGPT

Company line:4.8:1 marginal

Strengthened gradient (black/80→/90) and bumped text opacities: company line clears the 4.5:1 AA floor comfortably.

components/Projects.tsx
@@ card interior @@- <div className="absolute inset-0 bg-gradient-to-t- from-black/80 via-black/30 to-transparent" />+ <div className="absolute inset-0 bg-gradient-to-t+ from-black/90 via-black/40 to-transparent" /> <div className="absolute inset-x-0 bottom-0 p-6">- <span className="... text-white/60">{project.category}</span>- <p className="... text-white/50">{project.company}</p>+ <span className="... text-white/75">{project.category}</span>+ <p className="... text-white/65">{project.company}</p> <h3 className="... text-white">{project.title}</h3> </div>
Asked
Design the work cards so the overlaid text stays legible across any image. WCAG AA is the standard, but optimise for the hardest real case: a bright, high-detail photo underneath.
Came back
The first pass passed AA on paper. The company name sat at 4.8 to 1, a hair above the 4.5 floor, and marginal is not done. I had the gradient deepened from 80 to 90 percent black at the base and the two secondary lines lifted, category to 75 percent white and company to 65. The title already cleared AAA. The point was not the numbers; it was that the worst image on the site should stop having a vote in the typography.
Kept
Every line over every project photo clears AA with room, so new images can be added without re-testing, and the hierarchy of title, category and company holds whatever is behind it.

Layoutcomponents/Projects.tsx

Order at a glance

Equal tiles say nothing about which work matters. Asymmetry sets the reading order.

Click any card

Remitly
UoP
MSCI
MS

Each card shares a layoutId with its expanded view: framer-motion morphs between states.

components/Projects.tsx
<ul className="grid grid-cols-12 gap-4"> {projects.map((project, i) => { const colClass = [ "col-span-12 sm:col-span-7", "col-span-12 sm:col-span-5", "col-span-12 sm:col-span-5", "col-span-12 sm:col-span-7", ][i] return ( <motion.li key={project.id} className={colClass} initial={{ opacity: 0, y: 32 }} whileInView={{ opacity: 1, y: 0 }} viewport={{ once: true }} > <motion.div layoutId={project.id} onClick={() => setSelectedId(project.id)} className="relative rounded-3xl overflow-hidden cursor-pointer border" whileHover={{ scale: 1.025 }} whileTap={{ scale: 0.975 }} /> </motion.li> ) })}</ul> <AnimatePresence> {selected && ( <motion.div layoutId={selected.id} className="fixed inset-0 z-50 flex items-center justify-center p-4"> {/* expanded project detail */} </motion.div> )}</AnimatePresence>
Asked
Build the projects section with a bento feel: different card sizes, and a click that expands a card into a full overlay with the project details.
Came back
Equal tiles were the obvious first answer and the wrong one, because four identical cards say nothing about which work matters. I asked for a real 12-column grid with alternating 7/5 splits so the composition itself carries a reading order. For the expand, the card and its overlay share one layout identity, so the element morphs between states instead of a modal appearing over a card. Backdrop blurs on open; Escape and click-outside both close.
Kept
An asymmetric grid that states hierarchy before the copy does, and an expand that is one element in two states rather than two elements pretending.
NextHow I work with AI

Go deeper

See how this thinking holds up at product scale.