The AI Stack Behind Ski Resort Empire
How one developer used multiple AI tools to build a game in just a few months.
When people learn that Ski Resort Empire was built with extensive AI assistance, the first question is usually:
"Which AI built it?"
The honest answer is: none of them.
And all of them.
Over the last few months I have used several AI tools throughout the project. Each had different strengths, different weaknesses and a very different role in the development process.
The most surprising lesson was not how much AI could do.
It was how quickly it became obvious that no single tool was good at everything.
The Experiment
Before starting Ski Resort Empire, I had spent a considerable amount of time learning about machine learning and generative AI.
I completed courses including the Machine Learning Specialization and Deep Learning Specialization from DeepLearning.AI, partly because I was curious about the technology itself and partly because I wanted to separate reality from marketing.
There is a big difference between hearing that AI can generate code and understanding how modern models actually work.
Eventually I wanted to move beyond theory.
I needed a real project.
Not a weekend prototype.
Not a tutorial.
Something large enough that I would discover where AI genuinely helps and where human expertise still matters.
That project became Ski Resort Empire.
Copilot: The Discussion Partner
Microsoft Copilot became my strategic discussion partner.
Most of its value did not come from writing code.
Instead, it helped me think.
Many design decisions started as conversations:
- How should special events work?
- Why do players stop playing?
- How can retention be improved?
- When should a reward be presented?
- Is a new mechanic exciting or merely complicated?
- What would motivate players to return tomorrow?
Sometimes the suggestions were useful.
Sometimes they were not.
The important part was having a system that could challenge assumptions, generate alternatives and provide a second perspective.
In many ways it felt less like a coding tool and more like a product discussion partner.
Claude Code: The Pair Programmer
Claude Code became the primary implementation tool.
The project started with AI-assisted code generation and quickly evolved into a collaboration.
Claude helped with:
- Project setup
- New features
- Refactorings
- Testing
- Build pipelines
- Development tools
- Static website generation
- Privacy policies
- Patch notes
- This very devlog
The biggest productivity gain was not writing individual methods.
It was handling large-scale changes.
Features that would normally require hours of repetitive work could often be implemented, reviewed and refined much more quickly.
That does not mean the code was accepted blindly.
Architecture decisions, code reviews, balancing decisions and final implementation choices remained my responsibility.
The role was closer to a pair programmer than an autonomous developer.
Gemini: The Artist
The area where AI had the most visible impact was graphics.
I am a software architect.
I am not an illustrator.
Without image generation, Ski Resort Empire would have looked very different.
Gemini helped create:
- Business icons
- Story illustrations
- Region artwork
- Promotional graphics
- Splash screens
Some images worked immediately.
Many did not.
A ski lift became a bench.
A restaurant looked like a Roman temple.
A helicopter skiing icon accidentally turned into a mountain rescue mission.
The process was less about generating images and more about directing them.
Learning to write clear visual prompts became almost as important as learning to write code.
Gemini was also used for generating the background music that appears in the game.
What Surprised Me Most
The biggest surprise was not the quality.
It was the speed of experimentation.
In the past, many ideas would have remained ideas because exploring them required too much effort.
Now it became possible to quickly create:
- A mockup
- A prototype
- An illustration
- A design discussion
- A first implementation
That did not guarantee success.
Most ideas still required iteration.
But the cost of testing an idea became dramatically lower.
What AI Could Not Do
The most important decisions never came from AI.
Questions such as:
- Is this feature actually fun?
- Is progression balanced?
- Does this reward feel satisfying?
- Will players understand this system?
- Is this mechanic worth keeping?
still required human judgment.
AI generated options.
Human experience selected between them.
Many ideas were implemented.
Many others were discarded.
The final responsibility always remained with the developer.
More Than a Technical Experiment
Looking back, Ski Resort Empire became more than a game project.
It became an exploration of what software development looks like when a single developer gains access to tools that would have required a small team only a few years ago.
The result is not a game built by AI.
It is a game built by a developer who learned how to use AI effectively.
The distinction matters.
Because the most valuable resource was never the model.
It was still the ability to decide what should be built in the first place.