AI First Game Development: Levelling Up My AI Workflow with Game #2

Check out Release The Beast on the App Store.
For those new to this series, I recently chartered my journey building a game using an AI-first game Development Approach. The first part of the series set out to answer the following questions:
Can AI help a rusty ex-coder with a passion for games, not a lot of free time with work and family, create a decent computer game? And if so, how much load did the AI take, how has the software engineering process changed since I was in the driver’s seat, and what other aspects of game design and development can AI help out with beyond writing code? And finally, what are the best tools out there to help me with this job?
If you want to catch up, here is the first post in the series, Part 0: AI First Game Development (Intro).
Re-ignition
After building and publishing Game #1, Bubble Pop Hero—Animals, I realized that gains using AI were far less than originally expected.
My expectation was around a 50–60% increase in productivity, but it probably ended up being around 10-20%. This reflects the mood of the time where the AI-Hype is settling and we’re venturing beyond the wizards’ curtain and trying to figure out how to crank the levers in the right way to create real impact and value.
The amount of work that went into Bubble Pop Hero — Animals was a significant investment outside of working and family hours, which led to a hiatus from starting the next project.
But you can’t keep a closet games developer down for too long. With some enforced post-operation recovery time, I bit the bullet and launched into the next project, Release the Beast.
Release the Beast
My collaboration partner, Daz, came up with the idea for this game. It was a riff on the previous game, which involved freeing cute animals by popping bubbles that contained them.
The concept of Release the Beast is that a player shoots a puck and tries to get it to land on top of a spherical prison where the Beast is trapped, thereby freeing it. Think Shuffleboard meets mini-golf. Freed beasts are then added to a Sticker book. Each level has a unique and charismatic beast to free and different obstacles to navigate.
Daz cooked up a ton of designs for the game, which was a great starting point for launching into development—following an AI-first approach, of course.
This was a new opportunity to level up my workflow and try to get more bang for my buck from the latest AI Tools.
AI Tooling
My chosen tools for this project were ChatGPT and Claude for general-purpose Q&A, content generation and (in some cases) code generation. I did experiment with the new version of ChatGPT o1-preview and o1-mini.
I didn’t use AI image generators at all as Daz is a graphic and illustration powerhouse and churned out everything we needed.
The biggest change in my workflow was getting deep into AI Integrated Developer Environment (IDE) extensions, namely GitHub Copilot. I also experimented with the new AI IDE on block, Cursor.
Here is a breakdown of my experiences and learnings.
ChatGPT o1
At the time of writing, there was much hype about the new ChatGPT Model, o1, which offered more advanced reasoning capabilities through Chain-of-Thought techniques.
With o1 you make a request and it *thinks* for some time before responding. The thinking steps are communicated via the UI to show you what a clever little robot it is being.
In my testing, I had a cynical feeling—despite knowing how well O1 performed on benchmark tests and how well it was received in the tech community—that this was, to some degree, smart OpenAI marketing.
I tested 01 on complex challenges, and in some cases, it performed better than Claude or ChatGPT 4o, but at other times, it didn’t. I ran out of tokens pretty quickly then switched to o1-mini, then back to the old staples.
The Verdict:
Beyond being initially sold by the thinking steps the output was a mixed bag in terms of solving my problems. It wasn’t consistently more brilliant making me desperate to commit fully to this new model. This is almost certainly due to my use cases which were not that complex, so much more experimentation is required to road-test the next generation of Reasoning LLMs.
Cursor vs GitHub CoPilot
The biggest change in building Release the Beast was getting stuck into GenAI development tools available through Visual Studio Code. I started off with GitHub CoPilot and began tabbing my way through code generation, as well as writing inline prompts/instructions in my coding panel and using the chat panel for more broad queries and requests. It was satisfying to shift away from jumping back and forth between a chat client (Claude, ChatGPT) and my IDE.
During development, I heard about Cursor, a new AI-IDE on the block. There was a lot of excitement, and religious debates raged online about the virtues of each product.
Cursor intrigued me. The founders’ motivation to build a new product (facing up to the Goliath Microsoft) was that product development at Microsoft would always be bogged down by bureaucracy and couldn’t innovate in a way or at the pace the team envisaged.
Cursor is a fork of VS Code with custom GenAI features, many of them similar to GitHub Copilot. However, some attractive additions beckoned to the early adopter in me.
The Verdict:
Unfortunately, I had to switch away from Cursor to Copilot as it wasn’t playing ball with my Unity project. (Some features stopped working and became out of sync with my code)
However, my first impressions were positive. It used Claude (my preferred model for code) and had some features that Copilot didn’t have, like:
- Chat with code using RAG;
- Multi-line code complete;
- A feature called composer that can create entire applications or additions inside your project – including scanning your code to follow your style and creating the requisite files;
- and being able to make changes across multiple files.
(Aside: GitHub is hot on the heels of these features so it will be an interesting race to follow).
First impression: Cursor wins due to cool features.
Avoiding the AI-Brain Drain
Something I became hyper-conscious of during my development was the line between getting AI to help and getting AI to do everything.
Using AI as an aid can be invaluable in helping you learn and develop, but it is also easy to get lazy and just offload more and more to AI, which then starts to inhibit your growth.
It’s very easy to step over this line, and you can get into a loop of asking AI for code, scanning it, testing it, and then asking it to fix the mistakes without engaging deeply in the problem.
This behaviour extrapolated to general AI use e.g students writing essays, is dangerous and poses a real threat to cognitive development.
Educators of the world will undoubtedly wrestle with this and need to devise ways of forcing learners to battle through the nitty-gritty-painful complexity to build their mental muscles.
But I’m getting ahead of myself.
The verdict:
If you are producing something complex that you need to understand not only for the immediate challenge but for future and similar challenges, don’t let AI be a crutch. Let it be a passive guide while you own the direction and the execution of what you are doing.
What’s Next?
With Release the Beast up on the App Store, the next step is to get it on other platforms, market it and wait to get rich while the masses become addicted.
A bulletproof plan.
My biggest wish at this stage is twofold.
Firstly, having an AI agent that can do social media marketing. I am terrible at this and don’t enjoy it in the slightest.
Secondly, to have AI take care of the publishing to new platforms process which is always onerous, requires back and forth with the approval bodies, and fixing some weird and annoying bugs. I enjoy this even less than social media marketing.
Hopefully the next post I will celebrate using AI taking on the jobs we hate!