Over the past 2 months I built and fully shipped two mobile 3D games almost entirely with Claude Code.
I am senior web/mobile full-stack dev and have more than 15 years of experience, worked on countless apps, websites &amp; some 2D games (But never 3D games!).
Block Orbit
<img src="https://media.daily.dev/image/upload/s--3QmsHKiY--/f_auto/v1773991300/ugc/content_c125df6b-0df4-4702-ad04-9220afd84def?_a=BAMAMiiu0" alt="IMG_3034.png">
<img src="https://media.daily.dev/image/upload/s--kXOJ_0-X--/f_auto/v1773990742/ugc/content_13c6ad4a-7dbe-4328-bff5-0a0fe6fb1aa1?_a=BAMAMiiu0" alt="IMG_3023.png">
A puzzle game where you place block pieces onto a rotating 3D cylinder. Think Block Blast but wrapped around a cylinder so the columns connect seamlessly. Metal rendering with HDR bloom, particle effects, and every single sound in the game is synthesized in real-time with no audio files. 100 adventure levels across 10 worlds.
Built with Swift, raw Metal 3, procedural audio via AVAudioEngine.
<a href="https://apps.apple.com/us/app/block-orbit-3d-stack-match/id6760600736" target="_blank" rel="noopener nofollow">App Store</a>
Gridrise
<img src="https://media.daily.dev/image/upload/s--dcKRVho7--/f_auto/v1773990791/ugc/content_1427cf79-961d-46bc-8a65-5f72619220cd?_a=BAMAMiiu0" alt="IMG_2888.png">
Sudoku-like Square puzzle where the numbers are replaced by 3D Colored Towers. The twist is that you must deduce where to place the towers based on what is visible from the edges of the board. I later learned there is a game like this already called skyscrapers!
Built with React native, Expo, React Three Fiber (R3F), Three.js
<a href="https://apps.apple.com/us/app/gridrise-skyscraper-puzzle/id6759220447" target="_blank" rel="noopener nofollow">App Store</a>
<a href="https://play.google.com/store/apps/details?id=com.omar.gridly" target="_blank" rel="noopener nofollow">PlayStore</a>
<h4>What worked well</h4>
The speed is the obvious one and it’s extremely hard to overstate. Features that would normally take me a full day were done in an hour. All the logic, mechanics, the entire UI, Game Center integration, partner SDK setup, level parsing, save systems. Claude just ate through it.
Ideation is also fast and fun, brainstorming with Claude and then having it prototype and iterate without leaving the browser is really nice.
Repetitive mundane and tedious publishing related tasks:
<ul>
<li>Creating 30+ achievements (each with a unique icon, description and game design config)</li>
<li>Creating screenshots, promo-material and descriptions for App stores.</li>
</ul>
The two things above are probably the main reasons why I did not publish as many games pre-AI.
I enjoy the game-design and coding part, but the former mentioned tasks are very boring and tedious for me.
That’s when <a href="https://code.claude.com/docs/en/skills" target="_blank" rel="noopener nofollow">Claude Skills</a> came to the rescue.
For the above 2 issues, I used these 2 skills:
<ol>
<li><code>/generate-image</code> I asked Claude to create a script to use my Gemini API Token and use nano-banana image generation API to create a skill that allows Claude to generate images, I would then use it like this:</li>
</ol>
<code>check @achievements.json file, for each item there, use /generate-image to create an icon, generate all the icons in a square aspect with a dark blue background, the icon itself should be contained in a circle, use @ref.png as the base</code>
<img src="https://media.daily.dev/image/upload/s--z4nqEbuY--/f_auto/v1773990853/ugc/content_daa7f09e-30ae-4334-b73b-4a87f961f0c6?_a=BAMAMiiu0" alt="image.png">
What is cool about this technique is that Claude will create a unique prompt for each image generation request, and it will inspect each generated image based on my requirements (as outlined in the skill definition), if the generated image does not satisfy the requirements, he would then try again until the Gemini API gets it right.
<ol>
<li><code>/app-store-screenshots</code> <a href="https://github.com/ParthJadhav/app-store-screenshots" target="_blank" rel="noopener nofollow">(Source)</a> A really cool skill that generates App Store screenshots based on a simple prompt. I just had to provide the game name, a short description and some screenshots, and it generated 5 unique screenshots with different layouts and styles. It even added text and UI elements to make them look professional. What is really impressive is that it scaffolds a full Next.js project with all the code to generate the screenshots, so you can easily customize it or run it locally if you want to. OOB it did not support iPad screenshots, but I just had to ask it to add that feature and it did it for me.</li>
</ol>
<img src="https://media.daily.dev/image/upload/s--TQHR1bhT--/f_auto/v1773990884/ugc/content_34af88ed-291a-4a0b-b545-821578146b69?_a=BAMAMiiu0" alt="image.png">
<ul>
<li>Other parts that were very intimidating and were completely unknown to me were things like 3D Geometry and shader code. Claude wrote Metal/Three.js shaders (vertex, fragment, bloom, gaussian blur, tone mapping). given my lack of experience here I did not have high expectations, it did take a lot of iteration though, but I am still happy with the result.</li>
<li>Iterating on game-feel through conversation is also way faster than doing it manually. I could say “the ghost piece should pulse red when invalid” or “add magnetic snap when dragging near an invalid position” and get exactly what I meant (most of the time), I noticed that being descriptive and having command of language is very important, prompts like “make it really pretty” often lead to bad results.</li>
</ul>
What was harder than expected
<ul>
<li>
You still need to know what you want. Claude doesn’t design your game for you (yet at least). If you don’t have a clear vision you’ll get generic output, if I am feeling tired or lazy and just ask for “a cool shader effect when you place a piece” I might get something that is not what I want at all, and then I have to iterate on it wasting so much time (and tokens!).
</li>
<li>
Context management on a large codebase requires effort. I maintained a detailed <a href="http://CLAUDE.md" target="_blank" rel="noopener nofollow">CLAUDE.md</a> with the full architecture and several .md files that had (game-design) specifics. Without that it would constantly lose track of how things connect.
</li>
<li>
Debugging rendering issues is rough. When a shader produces wrong output Claude can reason about it but can’t see what’s on screen. You end up describing visual bugs in words which is slow and awkward. And it does occasionally introduce subtle bugs while fixing other things. You have to actually review the code. It’s not something you can just let run unsupervised.
</li>
</ul>
I have no monetary goals for these projects, I enjoy thinking about game design and making games, and AI is really making the hard and annoying parts easier, it is no silver-bullet though.
All worthwhile tools have a sharp edge that could cut, and needs to be handled with care!

Omar Alhaddad

A senior developer with 15+ years of experience shares a postmortem on building and shipping two mobile 3D games almost entirely using Claude Code. The games — Block Orbit (Swift, Metal 3) and Gridrise (React Native, Three.js) — were completed in two months. Key wins include dramatically faster feature development, AI-assisted brainstorming, and automating tedious publishing tasks like generating achievement icons and App Store screenshots using Claude Skills. Challenges included the need for a clear design vision, managing context on large codebases via CLAUDE.md files, and difficulty debugging visual/shader issues since Claude cannot see the screen. The author concludes AI is a powerful accelerator but not a silver bullet.

Gamedev with Claude Code - A retrospective