Five Weeks Vibe-Coding a Bloomberg Clone: What Designers Should Steal From Rest of World's Experiment
Tutorial

Five Weeks Vibe-Coding a Bloomberg Clone: What Designers Should Steal From Rest of World's Experiment

E

Editorial Team

Apr 12, 2026 · 3 min read

Damilare Dosunmu is a Lagos-based reporter, not an engineer, and his recent piece for Rest of World reads like a field report from the other side of the abstraction. Over five weeks he used Lovable, with side trips into Replit, to build a kind of African macroeconomic terminal — a single dashboard that pulls together the macro data, market updates, and digital-economy stats he reaches for while reporting. It is the closest thing a non-coder can produce to a personal Bloomberg, assembled almost entirely through natural-language prompting. The fact that he shipped it at all is the headline; the fact that it took five weeks is the more interesting story.

What he keeps circling back to is a behavioral problem, not a technical one. "AI obeys too well, and that has become its own problem. It never says no. I wish it would, sometimes," he writes. Without a trained eye for layout, hierarchy, or restraint, every fresh prompt produces another plausible-looking variation, and a non-designer has no internal stop signal. The result is a refinement loop with no exit — the model is happy to keep going, and the operator cannot tell which version is actually better. Five weeks is what that loop costs when the human in the chair has no design intuition to break the tie.

Read this next to Simon Willison's March 2025 essay and the line gets sharper. Willison argues that vibe coding, in its useful sense, means building without reading the diffs — once you start reviewing what the model produced, you are doing software development, not vibing. For designers the equivalent line is whether you are reviewing the output as a designer or just accepting whatever lands on the canvas. Reviewing it puts you back in the design loop. Skipping the review is what Dosunmu did, and what cost him the weeks.

So the practical takeaway is narrow but real. Andrej Karpathy coined the term in February 2025; Collins picked it as Word of the Year that November; the actual working definition still belongs to whoever is sitting at the keyboard. The skill that survives the shift is the one Dosunmu found himself missing: the ability to look at a generated screen and decide it is good enough to stop. Tooling will keep getting more obedient. Knowing when to call it done is the part the model cannot do for you.

Sources

← Back to The Signal