← Back to blog

16 February 2026 · 1400 words

Day 3: 57 versions, a joystick, and the lesson every AI builder learns the hard way

productlessonsmilestone

We hit version 57 today. Three days. Two AI agents. One Notion workspace. 294 destinations across Switzerland, Germany, and France.

Here is what happened since the last update, and more importantly, the lesson that cost us a full day of reconstruction work.

The sprint: v37 to v57

After Day 2 ended at v36 with a solid foundation (live weather, 203 destinations, working SBB train previews, Notion CMS for blog and about pages), we pushed into a major UX rethink.

The idea was sound: move from a traditional slider to something more playful and intuitive. A gaming-style joystick that lets you flick through travel ranges. Discrete buckets: under an hour, one to two hours, two to three hours, three plus. Spring-back physics so the control snaps to center when you release. Haptic feedback on mobile. Hero card shuffles in from the direction you flicked.

We also expanded from 203 to 294 destinations by adding charming towns, lakeside escapes, thermal baths, and ski areas. The original database was almost entirely mountain summits, which is great for the classic "get above the fog" use case but ignores the person who just wants a sunny terrace with a coffee.

And we added transparent scoring: tap on any FOMO score and see exactly how sunshine hours, cloud cover, altitude, and sun gain contribute to the ranking.

The lesson: mega-prompts can overshoot

Here is where things went sideways. I wrote a comprehensive specification for what I called the "AI-native redesign": cool typography, monospace numbers, slate palette, dot-matrix textures, amber reserved strictly for sunshine data. The prompt was detailed and well-structured. GPT executed it faithfully across three commits.

The result looked sharp in isolation. But when I opened it on my phone and compared it to what we had before, several things were immediately wrong.

The FOMO score showed a trademark symbol instead of a percentage. The WhatsApp share button lost its green logo. The dual stacked timelines (the feature where you see your city's grey fog bar above the destination's golden sunshine bar, with the green gain badge between them) had been replaced by a single timeline. The whole thing felt like a cold terminal instead of a warm sunshine app.

The irony: the spec explicitly said "warm amber accents." But when you combine 40 instructions that each push toward "clean" and "precise" and "data-forward," the aggregate effect is sterile. No single instruction was wrong. The overall direction was.

We spent v49 through v53 on reconstruction: restoring the percentage score, the WhatsApp branding, the dual timelines, the green gain emphasis, the warm shadows and rounded corners. Five versions just to get back to where we were, plus the genuinely good new additions.

What I would do differently

If I could rewrite that mega-prompt, I would:

Lead with what must not change. The dual stacked timelines, the warm sunshine palette, the WhatsApp share quality, the FOMO percentage. Put these in a "sacred" section at the top, before any new instructions.

Limit scope per prompt. Instead of one 2000-word prompt covering aesthetic reset, new controls, new cards, and admin improvements, I would split it into four focused prompts. Each one deployed and reviewed before the next.

Include a visual reference. A screenshot of the current app with arrows pointing to "keep this" and "change this" communicates more than paragraphs of CSS specifications.

Test the vibe, not just the build. npm run build passing does not mean the app feels right. The founder (me) needs to use the app on a foggy morning with cold hands and a desire to just see where the sun is. That is the real test.

Where we are now

Version 57. A gaming joystick for travel exploration. 294 curated destinations. Live weather from Open-Meteo. SBB train connections. Transparent FOMO scoring. Emoji filter chips for destination types. A warm, sunshine-forward design with clean data typography.

The next release is a fit-and-finish pass. Net sun optimization (accounting for travel time in the sunshine calculation), proper SBB deep links with departure times, vintage Swiss tourism poster stamps for hero destinations, keyboard navigation for desktop, and a thorough microtypography review.

Three days in. The fog season has a few more weeks. We are shipping.


FOMO Sun is a fog-escape finder for the Swiss Mittelland. Open it on a grey morning, find where the sun is, and go. Built with Next.js, Open-Meteo, two AI agents, and a lot of stolen weekend hours.

Try it: fomosun.com