How I built Whereto using AI

Start with a belief, then make its rule unbreakable

I opened a conversation, not an editor. I stated my belief, that a short list from someone with taste beats an endless algorithmic one, and asked Claude to break it: who is it for, who already does it, how does it make money, what is the smallest version that proves it? One sitting produced a brand doc, product requirements, a schema, a pitch and a six-week plan. The name changed later. The belief did not.

A belief hides a rule. Mine: ten picks, no more. So the database enforces it. Ranks run 1 to 10, no two picks share one, and nothing joins a list without something leaving. Every script Claude wrote after that had to respect it, because the database refused anything else.

Prototype cheaply. Let the evidence cut your scope.

The first plan had four categories (eat, drink, party, do) in a dark, austere iOS app. Reality subtracted. Landmarks and tours were hard to curate to the same standard, so they went, and party folded into drink. Within a day it was eat and drink only.

The look went from black to cobalt and butter yellow, and I went web first: a link can be texted in a second, a beta app cannot. Cities got the same treatment. If one was not good enough, I held it back rather than publish a thin list.

Design starts with taste, so borrow it from the best

AI will generate a design from one line. It will look fine and feel generic, and it cannot tell you which design is right. That is taste, and you build it by looking at a lot of good work.

So I collect references first. Refero is a searchable library of real, shipped screens: search "city guide" or "sign-in" and see how good teams solved it. Then I study sites I admire and name what I like: type pairing, spacing, photography, restraint.

Use references as direction, not templates. Pick three to five, write one sentence on what works in each, and give those to the AI with the screenshots. "Tight serif headline, quiet interface, one accent colour" beats "make it modern" every time.

Get a second opinion on your brand, then redesign with it

Once the site was live, I gave ChatGPT screenshots and asked for a brand analysis. A different model has not sat through your decisions, so it reacts like a new visitor. The key phrase is "keep what's working". Without it you get a teardown, not a refresh.

I pasted its refresh description into Claude Design and asked for a full redesign. The October 10 refresh brought a new typeface and a sharper line: "Know where to go." The analysis gave direction, the tool gave options, and my taste picked.

Choose boring technology and let the database be the bouncer

Keep the live parts to a minimum. Every page is a static file, so it cannot be slow or break. The database handles only accounts and the paywall, so locked picks never reach the page, even with JavaScript turned off.

Turn every bug into a check, and give the model ground truth

Every bug gets an automated check so it cannot come back. And when you know something for certain, like exactly where a place is, give it to the model as a hard constraint, not a hint.

Keep the voice human

Every pick has two lines: why it is on the list and what to order. They are the product, and the paywall protects them. Claude pulled raw material from real reviews; I wrote from there. The commit log is full of small corrections, including two city intros that a data change had quietly made false.

That is the sneaky failure: change the data and every sentence quoting it can go stale. Anything machine-written is labelled. Anything I stand behind, I wrote.

Write it down for the next session

Claude does not remember yesterday, so treat each session like onboarding a talented new coworker. I kept a README and an architecture doc in the repository, including the known weaknesses, which is the section a newcomer needs most. It also made an honest privacy page possible: written from what the code does, so the promise and the behaviour cannot drift apart.

Test with people you trust, then launch with a plan

Before launch I sent friends and family four asks: make an account, browse a city, use it when you are out, and tell me what I got wrong. A feedback form made that one tap. The fixes were unglamorous and important, and most were on phones. Next time I would test on a phone from day one.

For launch I asked for a schedule, not features. Each morning an email brings that day's post: the video, the caption, the hashtags, the time to post and a tracking link. I tap and paste.