Claude as Design Intelligence

Most people point AI at the keyboard. They ask it to write the code, fix the bug, ship the component, and they accept whatever comes back. I did something different. Before a single line of this site was written, I used Claude to make every decision the code would later obey, and then I argued with it. The design system, the color logic, the typography: all of it was reasoned out first, challenged, and only then handed to the build. Claude was not the hand that typed. It was the intelligence I pressure-tested every decision against.

That is the thread running through this entire series. The value was never in what the AI produced on the first try. It was in how hard I pushed on the second, third, and fourth. So that is the right place to start.

What Was Actually Wrong With the Old Site?

The site you are reading replaced a dark navy student portfolio. It looked fine. "Fine" was the problem. My audience is not other students. It is consulting partners and finance leaders between 35 and 65 who decide, in about eight seconds, whether I think clearly. Dark themes with gradients read like a SaaS landing page or a developer tool. They do not read like the firms I want to work for. Goldman, Apollo, McKinsey, Bridgewater: every one of them is light, quiet, and restrained.

So my first act was to challenge the request itself. I did not ask Claude to "redesign my site." I asked it to audit the one I had. A redesign is a guess. An audit is a diagnosis, and I wanted the diagnosis before I let anything be rebuilt.

How Do You Audit a Website With AI?

I ran the existing site through a design framework I built called ui-ux-pro-max, which turns Claude into a structured design critic rather than an agreeable assistant. It does not return vague encouragement. It returns a priority-tiered audit: every issue ranked 1 through 10 by leverage, tagged CRITICAL, HIGH, MEDIUM, or LOW, and sorted by domain across UX, style, typography, and color.

That structure exists precisely so I can argue back. When Claude tells you forty things are wrong, you freeze. When it ranks them, you can interrogate the ranking: I agree the 11-pixel body text is the top fix, but I disagree this color issue is HIGH, drop it to LOW. The audit was not a verdict I accepted. It was an opening position I negotiated.

How Often Did I Tell Claude It Was Wrong?

Constantly, and that was the job. The clearest example was the first light-theme pass. When you flip a dark site to light, the secondary surfaces still quietly point at the old dark accent, so half the page came back rendering as solid navy blocks. Claude's output technically did what I asked. It was still wrong. I rejected it and made it convert every card to white with a thin stone border, because "I made it light" and "it looks light" are not the same claim.

I built a gate to force these fights early. For any change touching more than two or three files, I made Claude work in plan mode first: articulate exactly what it intended to do before it was allowed to do it. Twice, on the two largest structural changes, it had misread me, and I caught the misinterpretation in a single message. Catching it after Claude had already rewritten twelve files would have cost a full rollback. To say the least, reading a plan is cheaper than undoing a build.

The pushback was not only technical. When the instinct was to prove range with logo banners and an "AI Lab" showcase, I argued the opposite: one disciplined Seton Hall navy mention, and a section called "Process" instead. Showcase says look at everything I touched. Process says I shipped a real product, here is how. The quieter choice was the more confident one, and I had to hold that line against the easy, decorated default.

What Does Institutional Restraint Actually Look Like?

What survived all that argument became the visual system. The palette is four colors, allocated on a 60-30-10 logic: Paper at 60 percent, ink at 30, slate at roughly 7, and a single deliberate touch of navy at 3 percent. Four colors as a hard ceiling, not a starting point. The type pairing is Source Serif 4 for headings and Inter for body, sized on a 1.25 modular scale stepping from a 60-pixel hero down to a 13-pixel label. No glassmorphism, no neon, no gradients. Restraint was not the absence of a design language. Restraint was the design language, and defending it meant rejecting most of what an AI eagerly offers.

Where Did the Build Actually Fight Back?

The hardest pushback was against bugs that hid. My body copy was rendering around 11 pixels because shadcn, the component library underneath the build, hardcodes a small text size you cannot reach from the outside. I refused the surface fix of stacking overrides and made Claude diagnose the cause, then build a custom card primitive with the correct 18-pixel size baked in at the source. The second fight cost a full step: Tailwind CSS v4 demands HSL color values, not hex, and feeds you nothing when you get it wrong. No error, just a palette silently ignored. Only by not accepting "it should work" did we find it.

What Does This Process Really Prove?

Anyone can generate a website now. That is no longer the signal. The signal is judgment: knowing when the first answer is wrong and being willing to say so. I used Claude the way a strong operator uses any sharp tool, to move faster and decide better, never to outsource the deciding. The decisions came first, the argument came second, the build came third, and the order is the entire story.

That same posture carries into Part 2, where the design system meets Lovable and React, and the question becomes how hard you push when the AI builds something that runs but is not yet right.