Genesis: Building This Site with AI Agents

I built this site in a weekend by directing AI agents instead of writing most of the code myself. This isn't a revolutionary workflow yet—it's an experiment in how much I can focus on architecture while agents handle implementation.

I used Claude for complex logic and structural decisions, and Gemini for design feedback and UI refinement. The goal was to see if I could stay in "architect mode" and skip the usual back-and-forth with syntax.

---

How It Worked

The key was treating AI like junior developers who need clear specs. I couldn't just prompt and hope. I had to:

1. Document first: I wrote out the vision, audience, and technical constraints before any code. This documentation became the shared context for every agent interaction.
2. Review everything: The AI wrote code, but I reviewed every file. I caught hallucinations, refactored unclear logic, and made architectural calls.
3. Use models for their strengths: Claude handled the Next.js structure and content pipeline. Gemini helped with design decisions and writing.

The bottleneck wasn't syntax anymore—it was whether I could articulate what I wanted clearly enough. When I was vague, the output was messy. When I was specific, it worked.

The Workflow

I acted as the architect and code reviewer. The AI acted as the implementer.

I started with a structured documentation phase: vision doc, audience doc, feature requirements, technical constraints. This gave the agents enough context to make reasonable decisions without constant correction.

For each feature, I'd describe what I wanted, review the generated code, and iterate. Sometimes the first pass was perfect. Sometimes it took three rounds to get the structure right.

Model assignments:

• Claude (Opus/Sonnet): Complex logic, architectural decisions, most of the implementation
• Gemini: UI polish, design feedback, content writing

The Tech Stack

Built on Next.js 16, React 19, and Tailwind CSS 4, deployed on Vercel.

I chose these tools for fast deployment and minimal infrastructure management. The focus was on building features, not configuring servers.

The Result

The site is live and functional. It has a custom MDX content pipeline, theme system, and responsive design. It works.

The speed gain was real: what usually takes me weeks of nights and weekends took two days. But the tradeoff is that I had to be more deliberate upfront—loose specs produced loose code.

What's Next

I built custom agents (@Researcher, @Planner, @Implementer) to orchestrate this workflow.

I'll document the agent setup and prompt patterns in the next log.