Contributing to App review backlog
I’ve been using AI tools at work almost every day for the past year. But until recently, I hadn’t tested their code development capabilities myself. Armed with my personal Claude Max subscription, I decided to see how far I could get — not with a proof of concept, but with actual apps I could publish to the App Store.
You might have seen posts from Elon Musk and Nikita Bier on X about App Store review times getting longer as more and more apps are being created. Here’s my story of contributing to that increase with my own app projects.
What I built
I had two criteria: the app had to be publishable on the Apple App Store as a full app, and it had to be useful for at least one person (including myself 😊 ). Two ideas fit the bill right away.
Word Wizard Max — a game about forming and playing with words and letters. My wife and I recently bought Scrabble, but we’re notoriously bad at it. I tried different vocabulary apps in the App Store, and most were either loaded with ads or locked behind paid purchases.
Clear Lines — my mom really likes a game called Color Lines from 1992. There have been many modern variants, but none of them actually worked well on her phone. I thought it would be a good challenge to build something she’d enjoy.
How I built them
The tech stack: React Native + Expo for cross-platform shipping, VS Code with Claude and Gemini plugins, and Claude Opus on the Max 5x subscription as the main coding agent (with Gemini as backup on the rare occasions Claude had issues).
Here’s my recipe for creating and refining an app with AI:
Give a high-level description of the idea — how it should work, what it's inspired by, and the tech stack you plan to use (in my case, Expo + React Native). Ask Claude to write a detailed spec: executive summary, tech stack, project architecture, key features for the POC, and additional feature ideas.
I found it most effective to have a full description of what I wanted the first version to do — rather than starting from a blank screen and adding functionality step by step. This way, you can think through architectural choices upfront: which modules need to exist and what the tech stack implications are.
With Claude Code in VS Code (Opus 4.6), ask it to create the project based on the outlined spec. In its current state, Claude is really powerful at executing with Expo — I had slightly less success with other stacks like Godot — and it can normally produce some version of even more complex ideas.
It takes a couple of attempts to get the application to fully run. My approach was testing the app myself and reporting all initial mistakes and runtime issues back to Claude.
In both cases, the AI-generated version had a lot of elements that didn't make sense out of the box — my answer for those was deleting them. Once I got to a clean base version, I started adding new features I had initially envisioned but that weren't properly executed in the first pass.
Ask Gemini to improve the graphical interface of the app by providing screenshots of the current version and ideas for how it should look. Ask it to create visual assets to be used in the app. For each asset, download the image and clean it up in GIMP — make the background transparent, remove visual artifacts, and so on. Sound effects were sourced separately with free licenses.
Go through each step of using the app and see how it can be improved. For example, in Word Wizard Max, I wanted the ability to rearrange letters used in the games — it wasn't produced in the initial POC. While I got it working after formulating a more detailed description, there were multiple bugs. What I found successful here was using Gemini CLI to test alternative implementations and see what works better. There were edge cases Claude alone couldn't resolve even at maximum effort — coupling it with another model's perspective was an effective workaround.
Once I had working versions, I tested them and incorporated feedback from friends. I submitted the versions I was happy with to App Store review — this turned out to be the longest step. The whole approval process took about 10 days and required multiple resubmissions. The feedback was mostly technical: my app requested permissions it didn't need, I needed to prepare additional documentation and walkthroughs, and so on.
Here’s the important part: I did not change a single line of code myself. I’m familiar with programming, but throughout development, every line was written by AI. Important caveat — both apps are fully offline, so security concerns are minimal.
How long did it take in the end? Accounting for iterations on design, features, and fixing bugs, it took me about 20 hours for Word Wizard Max and 10 hours for Clear Lines of my own time. Four weeks from start to finish, primarily driven by the app review process and user testing, which were hard to accelerate.
The total cost: $100 for Claude Max and $99 for the Apple Developer subscription. Gemini came free with my Google storage plan, and there were no other expenses — I already had a $6.50/year domain, and website hosting for app-related pages is free with Cloudflare Pages.
The result
Both apps are available on iOS (App Store) and Android (APK) — check them out on the projects page. I’m personally very happy with them and have spent more than a few hours playing as a user, not a developer. I also found my customers — my wife is now a dangerously good word game player, and my mom is happy with her new Lines game.
With limited prior knowledge of building apps end to end, the help of AI agents allowed me to get through the process effectively and learn a lot along the way. Whenever I was stuck and didn’t know how to approach something — for example, sourcing a sound effect — I’d open a new chat with Claude and just ask how to do it. It’s an incredible learning experience that no longer requires attending a formal course before getting started with execution.
My takeaways
AI progress is impressive — it goes beyond specs and PoCs, even if you’re not an engineer. There was not a single moment where Claude and Gemini couldn’t execute on what I asked. It feels like magic when you hand over a list of 10 features with detailed descriptions, and after thinking for a while, Claude comments: “These changes would require a significant new module — let me write it in one go”, then “Wrote 1,500 lines in file”, and finally “Compiles cleanly — done.” You launch it, and there are at least 10 glaring issues — but it works.
AI massively accelerates idea testing, but you still need a human in the loop. Even at maximum effort, the models make silly mistakes. When I asked Claude to create a tutorial walkthrough for a module, it created something — but the walkthrough didn’t actually make sense when you followed it step by step. When I asked it to curate a list of priority words for Word Wizard Max, the first suggestions were “DIE”, “GIN”, “GIG” (imagine these being the first words a kid sees in a game! 😨). But it’s incredibly powerful at converting a detailed vision in plain English into an end product — if you have a deeply thought-through design, Claude and Gemini can become your partners in crime.
Ideas are becoming worth less. The cost of building the first prototype is getting vanishingly small. If it takes 10 hours to go from an idea to a publishable app, having just a PowerPoint for a software-related idea might not be enough anymore. I’m not saying I’ll prototype every idea I have, but it looks like the world is heading in that direction.
What’s next
These two apps were an experiment to test what these tools are capable of. It was a great learning experience — from idea to navigating App Store rules to actually publishing. I’ll be on the lookout for interesting ideas to pursue next, and for ways to apply this approach in my work.