George Xing: how to run a one-person product team with AI agents

George Xing published this piece on Substack in May 2026. It is less a framework article than a technical diary — a direct account of how he structures his work as a solo developer building production software with AI assistance. The piece is organized around three systems that, in combination, let him match the output of a larger team.

The underlying insight runs against a common assumption: as code generation gets cheaper, the upstream work matters more, not less. Cheap code is only useful if you built the right thing. If the spec is vague, AI tools accelerate you toward the wrong outcome faster. Xing’s entire setup is designed around enforcing rigor before the model writes anything.

The first system is what he calls “agent skills” — a Claude plugin that bundles four sequential steps: brainstorming, plan writing, execution, and code review. Rather than chatting with an LLM and copying output, he is forced to complete each step before moving to the next. The code review step happens inside the same workflow, which catches issues before they accumulate into technical debt that is expensive to unwind later.

The second system is automated testing. Xing uses Codex for implementation review alongside custom testing integrations — XcodeBuildMCP for iOS and Playwright-MCP for web — that run user scenarios autonomously and flag failures. The effect is that he does not manually test most common paths; the system tests while he builds. This frees him to focus on edge cases and design judgment rather than regression coverage.

The third system is remote infrastructure: a Mac Mini as a persistent compute hub, Tailscale for networking, and SSH access that lets him continue work from any device. Synchronized Claude sessions carry context between desktop, mobile, and terminal environments, so context is not lost when he switches devices.

For product managers evaluating small-team AI development practices, the specific tools Xing uses are incidental. The transferable pattern is enforced structure upstream, automated validation downstream, and persistent context across devices and sessions. These properties scale regardless of team size, and they map directly to common pain points in AI-assisted development: vague specs, inadequate testing, and lost context between sessions.