Pairing with Claude Code to rebuild my startup's website (blog.nseldeib.com)

A non‑engineer founder rebuilt their startup CodeYam landing site by “pairing” with Anthropic’s Claude Code (via the Claude Code CLI) plus VS Code, GitHub CLI (invoked by Claude), and a Figma Dev Mode MCP server. Using a conventional developer workflow—local dev server (also previewable on mobile), feature branches, PRs, CI/CD and iterative testing—the author translated Figma designs into production‑ready code in weeks. Claude acted as both coder and PR reviewer, suggesting optimizations, renaming files semantically, and applying fixes; this enabled high‑fidelity implementation that low/no‑code builders couldn’t match while avoiding months of hand‑coding. The experience highlights both promise and limits of AI agents: Claude accelerated execution but exhibited variable response quality, could stop mid‑task, and sometimes made wrong or invasive changes (e.g., confusing files, generating unused hashed SVGs from the Figma MCP server). Practical mitigations included frequent commits/PRs, using Claude to search-and-clean hashed files, CI checks, GitHub Copilot for debugging failures, rollbacks and fresh sessions, and human developer oversight—especially for performance, accessibility, and architecture. For the AI/ML community this is a strong case study: agent-assisted development can compress delivery and augment non‑engineers, but safety, reliability, and workflow guardrails remain essential before trusting agents with core production code.