Claude × App Development for Non-Engineers ① - Creating, Testing, and Iterating. A New Approach to Product Development: "Growing" Your App

Using AI to build apps is no longer that unusual.

Until recently, AI had a strong image as an "excellent assistant" that helped with things like summarizing documents and brainstorming ideas. But lately, it's shifting into a "trusted partner" that actively leads development itself.

Our team members have been getting more opportunities to work with AI, and someone thought "why not try building something myself?" That's how this app development project started.

We were thinking "it would be great if we could build something useful for work," but the first idea that came to mind wasn't work-related—it was about a hobby.

Lately I've gotten into watching movies, and I've been seeing several films every week. I was keeping track of everything—what I wanted to watch and which films I liked—in my iPhone Notes app.

But my list was growing fast, and with so many similar titles, I kept wondering "wait, did I already watch this?" The Notes app management eventually fell apart.

So I decided to make my own custom tool, and that's the "Movie Watchlist" app I'm introducing today.

You can build a working app with rough prompts

When it comes to AI-powered development, there's an impression that "you need to write really detailed prompts that break down every requirement."

Many people believe that to create something that actually works, you need to organize the background and specifications first, then convey requirements without any gaps or omissions.

Of course, it's ideal if you can prepare and deliver the background and specifications all at once. However, what we learned from this experiment was that even with quite rough instructions, the AI understands our intent and implements it properly.

I think it also helped that it was a relatively simple application, but with just a prompt that threw together what we wanted to do, we ended up with a working app. The initial prompt looked something like this.

I want to create a movie watchlist.
For now it's just a prototype, so I'd like it to be contained in a single HTML file that runs immediately when opened in a browser.
The features should allow users to input the title of a film and where it's streaming (U-Next, Prime Video, etc.) to register it, and display registered films in a list.
Also, please add the ability to mark items as "watched."

With traditional development without AI, even the simplest project requires you to "design the UI," "select the technologies to use," and "write the code" — easily consuming several hours. Claude understood what we wanted in just minutes and quickly handled everything from optimal UI and tool selection through implementation.

The speed was so impressive that beyond admiration, I felt a touch of fear.

Once it's this fast, you naturally start wanting more. After the foundation was in place, we kept throwing every idea that came to mind at it. Those prompts weren't polished either — we just made bulleted notes of things we noticed and said "fix this," which is rough enough that any human engineer would definitely be annoyed.

After continuing this casual back-and-forth several times, the basic structure of the watchlist was completed easily.
Including fine-tuning, it took about an hour from the initial prompt to having a working app.

Think of it, implement it immediately. Grow it by adding as you go.

A few days after the framework was complete, the team had a session where each person introduced the app they created with AI, giving us a chance to see what other members had built. One of them had created an app with Gemini Nano integrated into it.

The moment I saw it, I thought to myself: "Come to think of it, I use Gemini to find movies I want to watch too. If I could build this into the app, I could search for movies based on my watchlist, couldn't I?" With that idea in mind, I headed home and immediately resumed development.

With a human colleague, suggesting scope creep mid-development like that would get me scolded—it'd be outrageous. But AI handles this kind of thing so smoothly. It implemented the movie search feature for me in just a few minutes.

So the search functionality came together without a hitch, but then I became concerned about something: accuracy.

At first, I kept the spec quite rough—recommending a certain number of films based only on the titles and genres registered in my watchlist. As a result, I often got movie suggestions that matched the genre but didn't align with my personal taste. I tried various approaches, like switching to a spec that picks films based on my favorited works, but the results didn't improve.

I realized that title and genre alone weren't enough information, so I decided to include plot summaries for each movie. But I absolutely didn't want to manually enter every synopsis myself—way too tedious. After wrestling with this problem, I decided to call on AI's help again.

I redesigned it so that plot summaries are automatically registered based on titles entered into the list. To ensure accuracy of the summaries, I had Claude write code that retrieves relevant information from the Wikipedia API based on registered titles and then uses AI to auto-generate the synopsis from that foundation.

Honestly, there are still some rough edges in the accuracy, and I'm considering an update to integrate with the TMDB API eventually. That said, we've reached a level where I can think of minor errors as "part of the charm." The entire fine-tuning process took only 2 to 3 hours.

As a side note: when I later tried to use the app on actual hardware, it turned out Gemini Nano wasn't available, so I switched to the Gemini API instead. While the free tier does have request limits, I could use a better model than with Gemini Nano, so accuracy improved significantly.

A quick look at the finished app

Building it as I went, literally "growing" the app with each tweak, I eventually realized it had come together with all these features.

Watchlist

• Register movie titles, genres, streaming services, and synopses
• Automatic synopsis generation using various APIs
• Toggle between unwatched and watched status
• Filtering by genre, favorites, and unwatched/watched status

Movie Search

• Gemini recommends movies based on your favorites
• Specify genre and number of recommendations
• Add recommended movies to your watchlist with one tap
• Add watched movies to your list with watched status

What used to be scattered between "searching for movies" and "recording movies" is now consolidated in a single app.
Previously, I had to explain my preferences to the AI each time I searched, but that hassle is gone. Now I can find movies I like anytime.

The watchlist—which was the original feature—is also now much easier to manage, since you can filter by viewing status, favorites, and genre, making it far more convenient than the notepad days.

What I personally appreciate is that when I find a work missing from the search results list, I can add it with just one tap.
This eliminates the hassle of having to manually type in the title and register it.

Summary

You may have noticed by now, but I haven't written a single line of code in this app development. I simply told Claude what features I wanted, and it implemented everything for me.

The first thing I realized when trying it out was that "I didn't need to decide on the final form from the start." As I used it, I'd notice things like "I wish this had that feature too" or "this isn't quite right," and I'd tell Claude about them each time. It felt less like "building" an app and more like "nurturing" one.

If you're hesitant because you think "I have to write the prompt carefully," just try throwing something together first. If the result doesn't match your image, you can refine it step by step—there's no such thing as real failure. Once you try it, you'll get a sense of "actually, this works" pretty quickly.

Please feel free to give it a try.