Building This Site with Agentic AI

Problem

A portfolio site is supposed to prove you can do the job you're applying for. For Mickey, that created a specific problem: he uses agentic AI as a real part of how he works — not as a novelty — and wanted the site itself to demonstrate that, not just claim it in a bullet point. But he'd never built a website before, had no local dev environment, and needed the result to be a real, professional, deployed site, not a toy.

The obvious failure mode was having an AI quietly build the whole thing with no visible process, which would prove nothing. The harder, more honest option was to actually collaborate through it — including the parts that didn't go smoothly — and document that.

Approach

The build ran spec-first: SITE_SPEC.md was written and agreed on before any code existed, covering positioning, site structure, tone, tech stack, and even a section on how the agentic build process itself should be documented. Every later decision traced back to it.

The site went up structurally complete before it was content-complete: all pages and routes were built on day one, with the unfinished case studies using resume language as honest placeholder copy rather than blocking launch on having every story finished. Content came later, case by case, each following the same pattern — Mickey supplying raw notes, a review pass to catch gaps or inconsistencies before drafting, then a custom Claude Code subagent turning reviewed notes into the actual page.

Not everything worked on the first try, and that friction is part of the story. The machine had no Node.js installed at all; the fix was using the Node already available inside WSL rather than a heavier native install. Three custom subagents were built for this project — one to draft case study content, one to handle deployment, one to review design — but the chat interface running most of this build turned out not to support invoking custom project agents by name at all, only a fixed built-in roster. Getting real named-agent invocation working meant installing the actual Claude Code CLI inside WSL, which surfaced its own snag: a stale system Node install shadowing the new one until nvm's default alias was fixed.

Deployment hit two real-world snags: Porkbun's default parking-page DNS records were still active and had to be replaced with Vercel's, and a GitHub privacy setting rejected a push over an email-address mismatch, resolved by switching to GitHub's provided no-reply address.

The site went live on Vercel before all the git workflow discipline was in place; a branches-and-PR convention, plus automatic post-merge branch cleanup, was adopted partway through once the production risk of pushing straight to master became real. GitHub's branch protection rules turned out to require a paid plan, so that discipline stayed a convention Mickey and Claude both follow, not something enforced by GitHub itself.

Outcome

A live, fully deployed site at mickwar.com with working custom-domain SSL, built and shipped in about two days of actual work.

All five other pieces of content on the site — four case studies plus the resume, toolbox, and about content — came out of this same process, and this write-up is the account of building the process itself, last, once there was something real to describe.

The friction points above aren't incidental — they're the actual evidence for the site's own thesis, that agentic AI is a genuine working habit rather than a polished marketing claim. A frictionless account would have been easier to write and less honest.