Stack And Bomb Development Process
What the hell can AI do for me?
What do you do when you purchase 30 days of access to Claude and the work you purchased it to do is completed in less than a day? After updating the Unity version to be store compliant for my mobile game Stack And Bomb, I was left with that question.
In 2018, the original game was released to the public in what could be described as a beta or even prototype state. The gameplay was solid, but menu was confusing and bland, the graphics were sub-par, monetization strategy was poor and it was missing something to drive engagement beyond a few levels. After 3 years of solo development, I was burnt out and just wanted to release it. I felt I had delivered my main objective of creating an original 2D arcade game.
I made the decision to use the remaining time on my one month subscription making the game better. How many close to my original intent could I make Stack And Bomb utilizing AI?
The poor graphics quality had been mentioned in reviews and feedback from friends. I set about to reskin many of the assets. I started with the easiest assets, the blocks. I wanted to retain the X, O block designs, but get away from white on gray.
Before Claude Design, I found improving graphics difficult
This was before I had access to Claude Design. I tried a Claude chat prompt and got terrible results. I have a Photoshop membership, so I tried Firefly. Firefly generated the most off topic random crap. I spent hours attempting to get something of use out of Firefly or Claude then manually adjusting in Photoshop. The process was annoying and time consuming, but that became the standard for all reskinning: First ask Claude to try, then fight with Firefly and finally spend hours manually cropping/blurring/adjusting in Photoshop.
It took me about one week to reskin the blocks, create new menu icon buttons, create scene backgrounds and update the foregrounds to be unique among the 6 scenes. I was very happy with the results. I like the outcome and I shaved off lots of time. I was super engaged experimenting with prompts and process. While I complain about manual work in Photoshop, overall I enjoyed creating the six themed stages.
To improve engagement I resurected original 2015 story concepts about a tech corrupted Squishyma Phundungous. Between levels were improved to show text that was a mixture of tips and Squishyma lure to define the universe in which Stack And Bomb exists. I utilized Claude Chat a lot to refine the history and the characters.
Screens between levels were improved to show the Squishyma still alive. I worked with Claude Code to devise a stable of uniquely colored and named characters, each with an assigned speed. When a player earned a new Squishyma for their team, the Squishyma would be introduced. In order to enhance the attachment a player would feel for their team, I added to the menu screens a way to view your current team and the name/age of each Squishyma.
Claude Code was a tremendous help in updating the menu system and my engagement enhancements. As long as I had clear directions for Claude Code, the results were fascinating. As an experienced developer, I am accustomed to many hours of coding and self QA. Claude Code needed lots of detailed instrutions and QA, but the coding was automated.
That is not to say that AI was a dream to use. You really need to stay involved, QA thoroughly and have patience. I use a Git Repo. I do not allow AI to commit or push to the repo. From my early days of working with computers I “save early and often”. Once I get something working, I commit and push the code. Version control allows me to let AI try odd things that I do not understand. Once or twice AI went in a totally wrong direction, and no amount of conversing with it would get it back on track. All I did was stash changes, kill the chat session and start with a new conversation with more clear prompting.
Claude Code occassionally assumed things about the existing code structure. Because Claude Code could not integrate with Unity at the time, I would allow it to write the code, but it had to instruct me on how to wire it up in the Unity interface. There were instances of Canvas Scaler orientation confusion, outline components not working with 9-slice sprites, etc… One full reset was when trying to adjust the scene camera Claude misread the portrait-versus-landscape and broke the camera system’s reference resolution settings.
Deploying to stores and monetization still sucks
Coding is fun. I love coding. How on earth did humans automate coding a video game, but setting up ads, creating in app purchases and deploying to app stores is still very much human torture?
For example:
On Android, Unity IAP 5.x ships no billing Library AARs - I had to manually add the Google Play Billing Library to both gradle templates, then deal with R8 stripping the calsses because they’re referenced via JNI reflection that the static analyzer can’t trace.
That is Claude telling me what I remember. I would have solved these problems in the old days Googling until I found an answer. I wouldn’t have cared what magic I used to fix an issue then, I don’t care now. That’s what the droid says happened, and I’m sticking to it.
