I’ve been building my portfolio site for a while now. Designed in Figma, built in WordPress with a fully custom theme, coded in VS Code with Claude Code integrated directly into my terminal.
It’s the most end-to-end design and development project I’ve taken on solo, and it’s given me a clearer picture than anything else of where AI genuinely pulls its weight, and where it absolutely does not.
This isn’t a theoretical take. This is what I’ve learned from actually doing it.
What the setup looks like
The workflow is three tools in conversation with each other: Figma holds the design system — variables, components, typography, spacing tokens, the whole architecture. VS Code is where the theme lives — PHP templates, CSS, JavaScript. Claude Code sits in the terminal, available when I need it.
The Figma-to-code pipeline is deliberate. Every decision made in the design system is supposed to show up correctly in the code. CSS custom properties map to semantic tokens. Component and layout structure in the browser mirrors component and layout structure in Figma. The goal is a codebase that reflects the design, not one that approximates it.
Keeping that fidelity intact while building alone, across design and development simultaneously, is where the AI question gets interesting.
Where AI wins
Catching errors I’d have stared at for an hour
This is the most unglamorous win and also the most real one. PHP is unforgiving. A missing comma in a wp_nav_menu() array, a malformed conditional, a function called before it’s defined — these are the kinds of errors that can take a disproportionate amount of time to find when you’re deep in a file and your eyes have stopped seeing what’s actually there.
Having Claude Code in the terminal means I can paste an error and get a precise diagnosis in seconds. Not a guess…a diagnosis. It reads the stack trace, identifies the line, explains what went wrong and why. What used to be a twenty minute debugging spiral is now a two minute fix.
Boilerplate and scaffolding
WordPress theme development involves a lot of repetitive structural code. Template files, function registrations, hook setups, menu registrations in functions.php — the code that needs to exist before anything interesting can happen. AI handles this scaffolding quickly and correctly, which means I get to the interesting parts faster.
Explaining the unfamiliar
I’m a designer who codes, not a developer.
There are corners of PHP and WordPress that I don’t know well, and rather than losing time to documentation rabbit holes, I can ask a direct question and get a direct answer. It’s made me faster at picking up new concepts and more confident making decisions in territory that’s less familiar.
Talking through problems
When something isn’t working and I’m not sure why, describing the problem out loud (even to an AI) forces enough clarity that the answer often surfaces in the process of articulating the question. This is the rubber duck debugging effect, except occasionally the duck has something useful to add.
Where the human wins
Design system coherence
This part is nuanced, and my thinking evolved as the build progressed.
Early on, I let AI take more of the lead on the development side than I should have, and what accumulated was a quiet mess. Hard-coded values were placed where variables should have been, and one-off decisions that didn’t know they were contradicting the system I’d built in Figma ran rampant. Nothing was broken…but everything was slightly wrong.
Here’s what I actually learned: AI isn’t inherently blind to your design system. It’s just blind to it without you. As I got more intentional about reinforcing the system — naming conventions, token structure, when to use a semantic versus a primitive — Claude Code started reflecting it back. By the end of the build, if I asked it to update link colors to match the blue theme, it would reach for --text--body--info rather than --blue--700. The right token, for the right reason.
The catch is that this doesn’t happen automatically. AI is a fast executor without a 100% persistent memory of the decisions that came before. So, the design system that lives in your head and is laid out all pristine in Figma is something that you need to consistently advocate for in every prompt. When you do that, the results get surprisingly good. When you don’t, entropy wins.
Visual judgment
AI cannot tell you when something is off. It cannot feel the difference between a layout that works and a layout that almost works. It cannot catch the spacing that is technically correct by the token scale but visually uncomfortable in context. It cannot notice that two shades of gray are close enough to create visual ambiguity without being close enough to read as intentional.
These are felt judgments developed over years of looking at design and developing taste. They are not computable. Every visual decision on this site is mine — not because AI couldn’t generate an opinion, but because AI’s opinion about visual design is not worth very much.
Understanding the product
My portfolio site exists to communicate who I am as a designer; my thinking, my process, my point of view, my taste. No AI has access to that. The decisions about what to show, what to say, how to structure the narrative of my work, what a potential client or recruiter needs to feel when they land on the page — those come from self-knowledge and strategic thinking that can’t be prompted into existence (well, as long as you’re trying to be authentic).
Understanding is always the human’s job. The tools serve the understanding. Never the other way around.
Knowing when something is wrong before you can explain why
This might be the most underrated design skill and the one AI is furthest from replicating. The instinct that fires before the analysis — the sense that something isn’t right, that a flow has a problem you haven’t located yet, that a component is going to break in a context you haven’t designed for. That instinct is pattern recognition built from years of experience. It is not something you can prompt for.
What this has actually taught me
Building this site has clarified something I suspected but can now say from experience: AI is a powerful amplifier on top of genuine expertise, and a liability in the hands of someone who hasn’t developed that expertise yet.
When I know what I’m doing, AI makes me faster. When I’ve been unclear about what I’m doing, AI has made the mess bigger, faster. The quality of the output is directly proportional to the quality of the thinking I bring into the conversation.
The portfolio site I’m finishing is mine. The design decisions are mine. The system architecture is mine. The thinking behind every template, every component, every content decision is mine. AI helped me build it faster and caught mistakes I would have otherwise lost time to. But it didn’t design anything. It executed, scaffolded, and debugged.
That division of labor — human thinks, AI executes — is the only version of this workflow I’d recommend. And only once you know your domain well enough to catch it when the execution goes wrong.
The question was never “AI vs human designers.” It was always: what does each do well, and are you honest enough about the answer to act accordingly?
