Sunset Call
A weekend tool that tells me if tonight's sunset is worth the drive, and how much to trust that call.
Role
Senior Product Designer
Team
Solo
Timeline
One weekend, October 10–11, 2026. Live since October 10
Status
Live at sunset.dimadidit.com
Context
Every afternoon I have the same question: is tonight's sunset worth the drive?
The Problem
A dull sky wastes an evening. A vivid one, if I skip it, is gone. Weather apps tell me about rain and temperature, not about color in the sky.
And a sunset forecast is never certain. So the real problem is not the forecast itself. It is trust: how do you show a recommendation so people trust it exactly as much as it deserves? That is also the hardest part of designing AI features, and a small tool is a good place to practice it.
Key Insights & Strategic Pivots
A clear sky is not a good sunset. Color needs a canvas. High clouds catch the light, and low clouds in the west block it.
Confidence without a track record is decoration. "7 in 10" is a promise. The only honest way to back it up is a record people can see, misses included.
The answer has to fit a glance. I check it on my phone, between other things.
Disagreement is information. When two weather models disagree, that is worth showing.
A tip you see every day stops being read. The advice to look west is the same every evening, so it should not take space every evening.
Strategic Goal
One screen that answers one question at a glance and shows exactly how much to trust the answer.
Design Strategy: six rules of trust
1. Answer first, then how sure
Answer first, then how sure
The Insight: In the afternoon I need a decision, not a weather report.
The Decision: The verdict comes first and large: "Likely vivid", "Maybe" or "Probably dull". Right under it: "Confidence about 7 in 10". The whole screen is a sky gradient, so the color tells the answer before I read it.
The Value: The decision takes a second, and the confidence is always one line away.
Why, one tap away
The Insight: A number alone asks for blind trust.
The Decision: "Why" opens three factors with real numbers and plain words: the air, high clouds ("the canvas") and low clouds to the west ("the light gets through"). One more line says whether the two weather models agree. The scale shows dull, maybe and vivid without a score, so a second number does not compete with "7 in 10".
The Value: I can check the reasoning, not only the result.
3. Check it yourself
The Insight: The forecast is a guess. My own eyes an hour before sunset are better.
The Decision: "Look west at 6:10 pm" opens what to look for: thin high clouds and a clear horizon mean go, a low gray wall on the horizon means skip. "Remind me" sends this check to the phone's Calendar as an event, so the page needs no notifications and no server.
The Value: The tool hands the final call back to the person, at the right moment.
4. Admit blind spots
The Insight: A confident guess two days out is worse than no guess.
The Decision: The further away the sunset, the lower the confidence. Beyond 36 hours the page says "Too early to say". When confidence drops under 4 in 10, there is no verdict, only "Hard to say" and the reason, for example "the models disagree".
The Value: Saying nothing where the forecast is weak makes the rest believable.
5. Keep score
The Insight: Trust has to be earned in public.
The Decision: After sunset the page asks "How was it?" with three answers: dull, okay and vivid. A call counts as right only on an exact match. Every miss stays on the timeline with a ring around its dot. Until five evenings are rated, the page says "Too few to judge".
The Value: "7 in 10" becomes something I can check, call by call.
6. Show freshness
The Insight: An old answer that looks new is a quiet lie.
The Decision: "Updated 4:18 pm · check again after 5 pm". When the data is old or the network fails: "Updated 11:05 am, 5 hours ago. Couldn't refresh." and a "Try again" button.
The Value: I always know how old the answer is.
Behind the Screen
1. The data
The page uses Open-Meteo with two weather models, GFS and ECMWF. One request covers three points: my location, and two points 50 and 150 km west along the line to the setting sun. The direction comes from the sun's azimuth on that day.
2. The formula
It is a transparent formula, not an AI model. High clouds plus half the mid clouds make the canvas, and around 40–60% is best. Low clouds along the line to the sun block the light. Full cover pushes the score close to zero. Rain earlier in the day and long visibility add a little, and haze takes a little away. The result is a score from 0 to 10. "How this works" explains it in five lines and names one known gap: a thin veil of high clouds can still glow, but here it scores low.
3. Average the opinions, not the sky
On the first afternoon, the two forecasts for the next evening disagreed hard. GFS saw 78% high clouds, ECMWF saw none. My first plan was to average the sky and score the result. That average sky scored 8.3, a confident "Likely vivid". Scored one by one, GFS gave 6.7 and ECMWF gave 2.0, and their average was 4.4, a "Maybe". Now the disagreement also lowers the confidence.
When two sources disagree, the design should show it and not blend it into a confident answer.
4. Honesty built into the code
Sample screens for demos use made-up weather and a "Sample" banner, and they never write to the history. A call that was not saved before sunset does not count. Every saved call keeps the version of the formula that made it. The location is rounded to about a kilometer, and the page explains why it needs the location before the phone asks for permission.
5. The build
Plain JavaScript, no backend, no accounts. The formula, the confidence rules and the score keeping are pure functions with tests on reference cases: a clear sky, full low cover, high clouds with a clear west. I built it with Claude Code as my coding partner and deployed it to my own server.
How I will know it works
"7 in 10" means I expect about three misses in ten evenings. The track record started on October 10, 2026. After 30 rated evenings I will check whether the confidence holds up. If it does not, the weights change and the formula becomes version 2, while old calls stay graded by version 1.
Final Thoughts
Sunset Call is a small tool. The rules behind it are the part I would take to any product where people act on a recommendation: the answer first, how sure it is, why, what to check yourself, where it is blind, how often it was right and how fresh it is. They work the same for a weather call and for an AI agent.
Let’s get in touch
I am ready to discuss this experience or apply it to new projects.