After a stretch of trying to build something the "right" way with disciplined, prompt-driven development, I took a break to indulge in pure vibecoding: an RSS reader styled after retro computer terminals and newswire printers. To lean all the way into the aesthetic I built it in Python, a language I barely know, which left me entirely at the mercy of the LLM.
The core came together fast — a terminal-like wall of headlines, and the vibes were good. Then the "one more thing" reflex kicked in. Claude had added a clock on its own, so why not local weather too? Selectable articles with descriptions? A more authentic animation when the feeds refresh? Each was almost effortless to add, and the project quietly grew more ambitious.
Then I noticed the count of new articles was wrong. Easy fix, I figured — a round or two of prompting. An hour later I was still at it: different models, the IDE, the web UI, each one confidently diagnosing a different "problem" and offering a fix that didn't work. I tried everything except reading the code.
What's a little embarrassing is what I eventually remembered: I never asked for that counter. Claude had added it, it looked useful, and I'd barely registered it. So I deleted it, and the bug went with it. When AI makes every feature feel free, the hard part isn't building. It's focusing on what matters and ruthlessly cutting what doesn't, even when "what doesn't" is work already done.
I ended up with exactly the nostalgic terminal news experience I wanted, and nothing I didn't. It's on GitHub.