Design Process

What does rigorous design actually look like when you can generate anything?

I built 40 versions of one portfolio in three days using AI. Most of them failed. Here is what the process actually looked like — the dead ends, the pivots, and the frameworks that emerged from the wreckage.

Kevin Nguyen · March 2026 · 10 min

Forty explorations in three days. The number sounds productive, even impressive, but most of those forty versions were discards — things built specifically to be abandoned. The real question the exercise surfaced is not how fast AI can generate designs, but whether the person holding the prompt has a rigorous enough process to know which ones to throw away, and why. That is the question this essay tries to answer.

When every visual idea is one prompt away, the bottleneck shifts. It is no longer the speed of rendering; it is the quality of judgment. A fast bad decision is just a faster bad outcome. The process I ended up with — breadth first, then structural diagnosis, then controlled subtraction — is less about AI and more about what a disciplined design loop looks like when the raw material becomes nearly free.

§ 01 — The premise

Start with the canvas, not the words

Before touching any portfolio content, I built four completely different visual effects. The goal was to explore the medium before committing to a message — to learn what the canvas could feel like before deciding what to say with it. Each effect was a different emotional register. The fluid simulation felt alive and responsive; it moved when you moved. The terrain effect felt slow and ancient, like something geological. The particle dolly felt cinematic but at arm's length — you watched it happen rather than participated. The reaction-diffusion growth felt alien, almost biological.

These are not features. They are moods. And knowing the difference between them — at the start, when all four are cheap to build — meant I could make a real choice instead of a default one.

§ 02 — Discovery

Testing interaction models before content

With the fluid effect chosen for its responsiveness and naturalness, the next question was how a visitor should engage with it. Three interaction models competed. One made scrolling change the physics of the fluid itself — scroll down and the colors warm, the movement grows more turbulent (E19, the Tide exploration). A second turned the cursor into a magnet that bent nearby elements. A third let you click to shatter and reassemble the layout.

The magnet effect was the most technically distinctive. It was also the most cognitively expensive: a cursor already busy bending things cannot also be a reading tool. The fluid won because dragging through it is natural — no one has to learn it. That is the test. When an interaction requires explanation, it has already failed at something.

How people interact with a design matters as much as how it looks. Choose the interaction model before you choose the visual system — they will constrain each other.

§ 03 — The pivot

Adding real content — and watching it break

This is where most design-with-AI projects fall apart. The visuals work. The interaction feels right. Then you drop in actual text and everything fights. E20 — Waypoints — was the first version with real portfolio content layered onto the fluid. The result was honest and instructive: content breaks everything.

Then came the review. I scored every variant on visual impact, content clarity, interaction quality, and performance. The verdict was blunt: beautiful noise is still noise. The content was present but not structured. A visitor could feel the craft but could not extract the signal.

The fix was not visual — it was structural. Darkening the background behind text made it readable but killed the fluid effect. No darkening meant the text competed with swirling colors. The real answer was to give content and visual effect their own separate space on the page — not to negotiate their coexistence on a shared surface.

When content breaks a design, the fix is usually structural, not cosmetic. Readability problems that look like contrast problems are often layout problems in disguise.

§ 04 — Convergence

Small changes, biggest differences

The instinct at this stage was to subtract, then reframe. I built four versions of the same design, changing only one variable at a time: fluid intensity, layout axis, transition style. E31 dialed down the motion — lower curl, lower splat force, higher dissipation — to feel calmer, more meditative. E36 added directional transitions: content sweeping in from the side rather than fading, giving the same underlying fluid effect a completely different personality.

The layout pivot came separately. Centering felt website-y — correct, but generic. Moving content to the left created asymmetry: the left half for reading, the right half for the fluid. Neither competed. Both breathed. This single layout shift solved readability, immersion, and hierarchy simultaneously. One structural decision, three problems gone.

§ 05 — The final form

Refinement is not polish — it is a new idea

E40 added the last layer: frosted-glass panels, spring animations where elements overshoot then settle into place, and a system I called Ghost Canvas. Ghost Canvas records every visitor's cursor movements and replays them as fluid disturbances later. When you visit the portfolio, the fluid you see is being shaped by someone who was there hours ago. The portfolio is never empty. It carries the trace of everyone who has been there.

The final version makes a deliberate hardware bet. It runs a high-resolution fluid simulation, frosted blur effects, and live network requests all at the same time. On an older laptop, you can feel it. The bet: the people reviewing this portfolio would have modern hardware. That is a reasonable assumption about the audience — and accepting it openly is the honest version of the trade-off, not the reckless one.

§ 06 — The loop in full

What the 40-version process actually looked like

Laid out end to end, the three days followed a pattern that is recognizable from product work: diverge quickly to understand the space, then find the structural diagnosis that unlocks the real problem, then converge through controlled subtraction.

§ 07 — The thesis

So, what does rigorous design look like when you can generate anything?

The premise of the question assumes that AI changes the hard part. It does not. Generating forty explorations in three days is fast because each generation is cheap. But the hard part — deciding which to keep, understanding why something does not work, finding the structural fix rather than the cosmetic one — that is still entirely human work. AI has not automated judgment. It has made the cost of being wrong very low, which is a different thing entirely.

What actually changed is the ratio between generation and evaluation. In a traditional design process, generating a new direction takes days; evaluating it takes hours. With AI, generating takes minutes; evaluating still takes hours. The bottleneck has moved. Process is now the constraint, not production.

AI has not automated judgment. It has made the cost of being wrong very low. The process question — how do you know which direction is right — becomes more important, not less.

The five frameworks that emerged from the forty explorations are not rules. They are heuristics for working at the new ratio — when generating is nearly free and discarding is the primary skill:

The process itself was a product exercise. Scope broadly. Test interaction models before visual systems. Integrate the real requirement — actual content — as early as possible. Refine through subtraction. Then ship something that reframes the problem. 40 explorations in 3 days — not because speed matters, but because momentum does. When the cost of a wrong turn is near zero, the only way to find the right path is to take many of them, fast, with enough judgment to know which ones to abandon.

Discarding is the primary skill. The portfolio exists because of the forty versions that did not ship — not in spite of them.

§ 08 — Open questions

What I would do differently

The hardware bet in E40 is the honest trade-off. A high-resolution fluid simulation, frosted blur, and live network requests running concurrently — on an older laptop, you feel it. The decision to accept that cost was made with a specific audience in mind: people reviewing a PM portfolio likely have modern hardware. That reasoning holds, but it should be stated clearly rather than absorbed quietly. Every performance trade-off is a product decision, and product decisions should be owned.

The second thing I would change is earlier content integration. I waited until E20 to drop in real portfolio text. Moving that test to E5 or E6 would have shortened the diverge phase and sharpened the structural diagnosis earlier. The lesson from the pivot — that content breaks everything — should be an input to Phase 01, not a discovery in Phase 02.