Building Systems with a Stroke of a Pen (alyx.substack.com)

An engineer recounts bootstrapping a cross‑platform marketplace by orchestrating a "team" of LLMs—treating models like different pens for sketching system architecture and code. Starting with ChatGPT 3.5 and moving through GPT‑4, Gemini, and Qwen, they parceled work into narrowly defined tasks, assigning models by strengths (large context windows, single‑pass problem solving, etc.). Key technical tactics include distilling vast context with Gemini 2.5 Pro’s 1‑million‑token window into concise overviews for smaller frontier models, letting models auto‑iterate by adding telemetry, and human-in-the-loop spot checks to catch overconfidence, memory entropy, and other LLM-specific bugs. Embedded IDEs like Cursor were useful only for autocomplete; complex, hands‑off code generation wasn’t trusted. The account highlights a shifting engineering abstraction: developers become architects and managers of coding agents rather than coding every line. That has practical implications—faster solo or small‑team development under tight burn constraints, new roles around prompt orchestration and model governance, and a higher baseline of systemic complexity. Crucially, the author argues humans must remain actively in control—guiding, validating and integrating LLM outputs—so AI augments rather than replaces engineering jobs.