This website—the one you are reading right now—was scaffolded using AI coding agents. Not as a gimmick or a case study, but because it was genuinely the most efficient way to get a professional, SEO-optimised site live while focusing my energy on the consulting work that actually pays the bills.

Here is what the process actually looked like, with no embellishment.

The Tools

  • Codex operates as an autonomous agent system. You assign it a task—"set up a Next.js project with MDX blog support, dynamic routing, and Schema markup"—and it works through the implementation in a sandboxed environment, presenting the result for review. In 2026, Codex has evolved beyond simple code generation into a full agentic workspace that can spin up environments, run tests, and prepare pull requests.
  • Antigravity is the AI coding assistant from Google DeepMind, designed for interactive pair programming. Where Codex excels at autonomous background tasks, Antigravity is built for the iterative, hands-on refinement—adjusting layouts, debugging edge cases, wiring up API integrations where you want to watch every change as it happens.

The Workflow

The honest vibe coding workflow was not a single magical prompt. It was iterative and highly conversational:

  1. Foundation via Codex: The initial project structure—file system, routing, configuration, content pipeline—was delegated to Codex as a clearly scoped task. This is the kind of boilerplate that an autonomous agent handles well because the requirements are well-defined and the output is verifiable.
  2. Refinement via Antigravity: The design system, responsive layout adjustments, SEO optimisation, and content formatting were handled interactively. These tasks require judgment calls and visual feedback, making them better suited to a real-time coding assistant.
  3. Content Pipeline: The blog's MDX pipeline (the system that converts markdown files into rendered pages with Schema markup) was built collaboratively—architectural decisions made through conversation, implementation handled by the AI.

What Worked

  • Speed: The foundation was functional within hours, not weeks.
  • Consistency: AI-generated code tends to follow patterns rigorously, producing consistent file structures and naming conventions.
  • SEO Compliance: Schema markup, meta tags, and semantic HTML were implemented systematically rather than retrofitted.

What Required Human Judgment

  • Design Taste: The AI can generate any layout you describe, but it doesn't have aesthetic opinions. The visual direction came from clear creative briefs, not from asking the AI "make it look good."
  • Brand Voice: The content strategy, tone, and editorial decisions are entirely human. The AI assists with structure and formatting, but the ideas and personality behind the writing are not delegable.
  • Integration Testing: Connecting to external APIs (Notion, Stripe, scheduling tools) required testing with real credentials and edge cases that only manual verification could validate.

The website you're reading exists because of a combination of human direction and AI execution. That is the "Human on the Balcony" in practice—not a theoretical framework, but the literal methodology behind this project.