The Control Tower and the Builder

I knew AI could write code for me. The question was: which AI, and how to direct it.

Two options. Claude Code, a terminal-based agent. Type into a black screen and AI creates files, runs code, fixes errors on its own. Or Cowork, a desktop app. Same AI but with a clickable GUI where you can drag and drop files.

Instinctively, Cowork looked easier. You can click things. The terminal still felt foreign. But before trusting instinct, I asked the AI itself: "Compare these two tools for this project. Three criteria—how much I need to be involved, quality of output, time to completion."

Asking AI to evaluate itself was a peculiar experience. The answer was candid.

Claude Code was overwhelmingly better suited. The reason was simple: in the terminal, AI can run the full cycle of writing code, executing it, reading errors, and fixing them—autonomously. "Fix until the tests pass" actually works. Cowork can create files but can't execute them. It can cook but can't taste.

So I split the structure.

The chat window became the "control tower." Overall design, progress tracking, error analysis. The terminal's Claude Code became the "builder." Executing the decisions made in the control tower.

This metaphor came naturally because my role became clear. I don't fly the plane. AI flies the plane (writes code). I sit in the control tower saying "move to runway 17" and "lower altitude." And I watch the radar (checklist) to confirm the plane follows its planned route.

Separating roles improved results. The control tower handles "why" and "what." The builder handles "how." Previously, I'd mixed everything in one place—asking about design, then immediately requesting code, then asking to fix errors in the same context. Output consistency suffered.

I'm the PM. AI is the capable junior. Does assigned work quickly and accurately. But can't independently set project direction or anticipate unasked-for needs. And occasionally builds features I didn't request—a junior's eager initiative. So the build prompt needed clauses like "do NOT write authentication code in this phase."

After work, I open the laptop. First the control tower. Check where yesterday ended. Decide today's task. Then open the terminal. Assign work to the builder.

Once this sequence became habit, limited time became an advantage. Only two or three evening hours means "today, only this" is enforced. Narrow scope makes AI instructions precise. Precise instructions produce precise results.

Less time means less room for mistakes. So people with only evenings become more strategic—something I learned around this time.

---

🔧 Technical Terms in This Episode

Terminal A black screen where you give computers text-based commands. Instead of clicking with a mouse, you type to create files, run programs, and start servers.

Agent An AI program that takes instructions, autonomously plans, executes, and verifies results. Unlike simple chatbots, it reads and writes files, runs commands, and self-corrects errors.

CLI (Command Line Interface) Text-based operation via terminal. The opposite is GUI (Graphical User Interface)—mouse-clicking. Claude Code is CLI-based; Cowork is GUI-based.

Debugging Finding and fixing bugs (errors) in code. From "catching bugs." In vibe coding, AI reads error logs and traces causes, making debugging possible even without reading code.