Building This Portfolio

AI is not magic. It does not replace thinking. You still have to articulate what you want, define what "good" looks like, and understand how the thing you are building actually works. Because you can build fast, you should think slow - slower than ever before. Every prompt should count. Every use should count. Review the code with your own eyes, understand what changed and why, and course-correct when it drifts. I am acutely aware of the environmental cost of these tools, which makes intentional, deliberate use a responsibility, not a nice-to-have - until we figure out how to make them sustainable for everyone.

With that in mind, here is how I designed a token system, built a component library, wrote the content, and shipped a production portfolio in a 48-hour sprint.

Before I opened Cursor or Claude, I had been collecting material for months. Case study metrics and slide decks organized in NotebookLM had been gathering for over a year. Font and color inspiration sites were in bookmarks but I forced myself to pick stuff and follow my gut with inspiration. I had a previous portfolio site with voice and tone I could riff on. A library of links, articles, and videos I had consumed about AI-augmented workflows. And most importantly, an idea I could articulate: a 48-hour portfolio sprint, tokens first, AI augmented, built as a beta test for a larger workflow presentation at work.

I also had some prior experience prompting and building with AI. Not deep engineering knowledge, but enough to understand what was being generated, ask good questions about it, and know when something was off.

Friday afternoon I sat down with Claude and started planning. My process is messy at first: I talk, ramble, type, refine, and iterate until the shape of the plan emerges. I gave it context about who I am, what I was building, what tools I had, and what I saw in my head. Over about two hours, I had a structured plan, a set of strong starting prompts, and a Cursor context file that would act as a guardrail for every agent interaction going forward.

Two hours sounds like a lot for "just planning," but I was re-reading plans, thinking through decisions, and pressure-testing ideas before writing a single line of code. Plans are where it all starts. If you are not planning, you are going to end up with slop.

Friday evening at 8:41, I started scaffolding in Cursor. The first prompt set up Next.js with my token system, extended Tailwind with all the token values, configured three Google Fonts, and created a utility helper. Then I had the agent build the full layout shell: container, header, footer, and a test hero so I could verify the type system.

The agent built the entire app skeleton with placeholders in one go. I had planned to work page by page, but the all-at-once approach actually worked fine. I did some manual tweaks to prompts along the way as we worked through the first few things. By midnight, after another 30 minutes finishing the layout shell and getting the case study data structure ready, I had the site running locally with themed skeletons of every page. Day one was done.

Saturday afternoon I hit a rhythm. Content refinement, visual tweaks, hybrid typing and prompting and committing. I got staging live on Vercel. Somewhere around 3pm I fell into a flow state and stopped documenting individual steps. I was just building: prompting specific changes, scanning the result, adjusting, committing. The key discipline was making sure we baked accessibility rules into everything as we went, not as an afterthought.

Saturday around 1pm, as I was editing case study content, I realized something: my case study content lived in the same file as my front-end UI tokens. That was the agent's suggestion, and I had run with it without thinking about it too much. But working in the file made the problem obvious. Content and design tokens are different concerns, and coupling them would create maintenance headaches.

I used Claude to talk through the separation plan in plain English, refined it together, and started the migration. Along the way we hit some routing errors. I spent time reading documentation, talking to the agent, and we got the fixes in. About 45 minutes total. It was a good reminder that you can course-correct once you have set a plan, and that verifying your instincts with AI and some searching builds real confidence in the decisions.

The other friction was knowing when to stop. During intense building, everything looks like it could be a little better. One more visual tweak. One more wording pass. The accessibility work helped ground me: running Lighthouse audits and checking focus states gave me concrete benchmarks instead of subjective "does this feel done?" It was clear what met the bar and what did not.

What AI got right by default through the token system was significant: contrast ratios, semantic heading hierarchy, and focus ring styles all came along for free because the tokens enforced them. What had to be explicitly enforced through human-in-the-loop review was subtler: reading order, link target clarity, and making sure nothing important was conveyed through color alone.

By Saturday night at 10pm, I was tired. I was playing around with Codex 5.3 so I had it summarize everything built that day and plan for Sunday, which was mostly getting my domain configured and the site live.

Sunday morning I woke up with a dream, and in that dream I had the idea to add a resume page to the portfolio. What if the resume was just a Markdown file that generated both a downloadable PDF and the website page? I took the idea into Claude, worked out a plan, and had it built in about two hours. I found a service that renders Markdown to PDF, so all I needed was the website version. After a few iterations making sure the Markdown looked good in both formats, I called it for the 48-hour sprint.

I spent another hour double-checking everything we built against the plans we had written, then updated all the plans with progress and changes made along the way. Then I deployed the site live, noticed a few tweaks here and there (as you always do), and shipped it.

Where it stands now: a living project with an iterative post-launch plan, custom agent skills I have built to help maintain the project, work on new ideas, and document changes easily. The pipeline from tokens to code to deploy is validated and repeatable.

Think smarter and work smarter, not harder. Do not slop it up.