← Back to list

Building This Site with AI: A Meta-Rambling

4 min read

This website you're reading was built entirely through conversation. Not with a team, but with Amp—an AI coding agent—guided by a specific prompting methodology that turned a weekend idea into a production site.

Since this is my first post here, it seems fitting to document the process. Not the technical stack (Next.js, Contentlayer, the usual suspects), but the collaboration patterns that made it work.

The Initial Prompt: Setting the Stage

I didn't start with "build me a website." I started with what I'd call an "architect prompt"—a structured brief that did three things:

  1. Established the vision: "Minimal personal website with a blog section called 'ramblings.' Think typography-first, content-focused design."

  2. Defined constraints: "Text-based, minimal but beautiful. No fancy features, just clean reading experience. Deploy to Vercel."

  3. Orchestrated the AI team: "Use Oracle for planning, spawn subagents for parallel implementation, iterate based on feedback."

That last point was crucial. By explicitly naming roles and workflows, I gave the AI system a clear organizational structure to follow.

Oracle: The Planning Specialist

The first tool Amp used was Oracle—a reasoning-focused AI designed for high-level planning and code review. My directive was simple:

"Consult Oracle for a comprehensive implementation plan. Break down the project into phases, identify potential issues, suggest the best technical approach."

Oracle delivered a detailed roadmap:

  • Phase 1: Project bootstrap and configuration
  • Phase 2: Typography system and global styles
  • Phase 3: Core components (Header, Footer, etc.)
  • Phase 4: Page routing and content management
  • Phase 5: Polish and optimization

This wasn't just a to-do list—it was a strategic framework that guided every subsequent decision.

Subagents: Parallel Execution

With the plan locked in, Amp spawned multiple specialized subagents to work in parallel:

  • Bootstrap Agent: Set up Next.js, Contentlayer, dependencies
  • Design Agent: Implemented typography, colors, layout system
  • Component Agent: Built reusable UI components
  • Content Agent: Set up markdown processing and routing

Each received focused, scoped prompts with clear inputs, outputs, and success criteria. This parallel execution turned what could have been hours of sequential work into concurrent streams.

The Iteration Dance

The magic happened in the feedback loop:

  1. Reference-driven design: I shared screenshots from stephango.com and other minimal sites as north stars
  2. Rapid prototyping: Subagents would implement, Amp would integrate
  3. Visual feedback: "This looks cramped" or "needs more whitespace" with specific examples
  4. Quick pivots: When something wasn't working, we'd backtrack and try a different approach

The key insight: giving visual references rather than abstract descriptions led to much better results. "Make it look like stephango.com" was infinitely more effective than "make it minimal."

What Went Right

Several patterns emerged that made this collaboration unusually smooth:

Clear hierarchy: Oracle → Amp → Subagents created a decision-making structure that prevented conflicts and confusion.

Scoped tasks: Instead of "build a website," each agent got narrow, specific objectives with measurable outcomes.

Iterative refinement: We built in layers—structure, then styling, then polish—rather than trying to perfect everything at once.

Reference-first feedback: Screenshots and examples gave the AI concrete targets to aim for.

The Human Element

What couldn't be automated was taste and judgment. The AI could implement any design I described, but knowing what to ask for remained distinctly human:

  • Recognizing when the initial design was too cluttered
  • Knowing when to stop tweaking and ship
  • Understanding the difference between "working" and "feeling right"

The AI was an incredibly capable execution partner, but the creative direction stayed firmly in human hands.

Lessons for AI-Assisted Development

If you're thinking about using AI for your own projects, here are the patterns that worked:

  1. Start with structure: Write a brief that covers vision, constraints, and team roles
  2. Use planning tools: Let a reasoning-focused AI create the roadmap before diving into implementation
  3. Scope narrowly: Give agents focused tasks rather than broad mandates
  4. Iterate visually: Share references, screenshots, and examples rather than abstract descriptions
  5. Know when to ship: Perfect is the enemy of done—stop when it's good enough

Meta-Observations

There's something fascinating about using AI to build a platform for human writing. The tools excel at the mechanical aspects—boilerplate, configuration, responsive layouts—freeing up mental cycles for the creative work that still requires human judgment.

This site represents about 3 hours of active prompting spread over a weekend. The result is clean, fast, and exactly what I envisioned. More importantly, the process felt collaborative rather than transactional—less like giving orders to a machine, more like working with a very capable (if literal-minded) partner.

The future of development might not be replacing programmers with AI, but rather augmenting human creativity with AI execution. At least, that's how it felt building this little corner of the internet.

Now, on to the actual ramblings.