I had never designed a game before. I did not know the terms, the common ideas, or the process. All I had was a simple wish to make something fun to play online with friends. That was enough to start a discussion with ChatGPT, and what followed was one of the most surprisingly productive creative experiences I have had.
This is the story of how that conversation developed, what the AI did well, where it made mistakes, and what it taught me about using artificial intelligence as a real creative partner.
Starting from Nothing — "I Want a Simple Horror Game"
My first prompt was broad and honest:
I have no experience in game design but I want to create a simple and lightweight online game which I can enjoy with my friends. The game should be survival horror.
I expected either a lot of complex terms or a basic reply. What I got instead was a clear, beginner-friendly plan.
ChatGPT quickly recognized the main challenge of keeping the project small and manageable. Instead of suggesting a big open world or complex systems, it focused on one simple gameplay loop: explore a haunted place, collect items, avoid a monster, and escape. One map. One monster. Four players. It even gave the idea a name: Night Shift.
The first game concept — core loop
Explore the map → Collect items → Avoid the monster → Complete objectives → Escape before time runs out
This was the AI doing something truly useful. It did not just give ideas, but also filtered them for someone who could not yet judge what was possible. It suggested Unity and Photon PUN for multiplayer, recommended low-poly graphics to keep the game simple, and pointed out that sound design, not visuals, is what makes horror feel scary.
For a beginner, this was the right kind of guidance.
A Twist - Bringing Mafia into the Horror
The conversation shifted when I started making the design more personal. I wanted the game to feel like a modern version of Mafia, the classic social deduction party game.
This is where the AI showed its ability to combine ideas. Instead of treating the two genres as separate, ChatGPT merged them into one clear concept.
"Think of it as Mafia plus exploration plus horror tension. Players move around the map instead of just talking in a lobby."
The result was a game where most players are survivors trying to escape a dark facility, while a hidden traitor works against them.
It also suggested active sabotage mechanics, such as turning off lights, breaking generators, and luring the monster. This gave the traitor role active and interesting tasks instead of just voting. This showed smart design thinking. It also used real examples like Among Us, Phasmophobia, Deceit, and Project Winter. The AI was not creating ideas from nothing. It was combining ideas from a genre I did not know about.
The Role System - Making Every Player Feel Different
The most detailed and useful part of the conversation came when I explained exactly how I wanted the roles to work. They would be random, include active tasks each round, and be designed so no one could easily guess your role just by watching you. After each night, someone would be killed or saved, and then the group would talk and vote out a player which will mark the end of a round.
ChatGPT turned this into a complete role system built around one elegant design principle: every player performs the same visible tasks, but the outcomes differ by role.
Lighting lanterns, fixing generators, delivering supplies — all players do these. But what happens as a result depends on who you are.
| Role | Visible action | Hidden effect |
|---|---|---|
| Villager | Repairs fences, gathers wood | None — relies on observation |
| Mafia | Pretends to complete tasks | Secretly poisons food, cuts wires, eliminates at night |
| Detective | "Inspects footprints" | Receives a hidden clue about another player |
| Doctor | "Prepares medicine" | Secretly marks one player as protected for the night |
The round structure
Each round would last about four minutes, which kept the pace fast and tense.
- Day phase — movement and visible tasks
- Night phase — hidden role actions resolve
- Morning phase — outcomes revealed
- Discussion & Voting — in town square
The AI understood my idea correctly. It recognized that hiding roles through shared tasks was the main concept, not just an extra feature.
The Standout Idea - When the Game Itself Lies to You
When I asked for one feature that could make the game truly unique, the AI gave its most creative suggestion of the whole conversation.
It proposed something called the Fear Reflection System — a hidden fear level that, when it gets high enough, makes the game world trick you.
You might see a player walk past who was never there. You might think you saw a kill, but nothing actually happened. You might hear someone call your name from an empty hallway.
This challenged a foundational rule of social deduction games: that what you see is true. It transformed the game from a social exercise into something more psychological and genuinely unsettling.
This was the most creative moment of the collaboration. It was an idea I would not have arrived at on my own.
Where the AI Stumbled - The Cost of Simplification
When I explained that visual hallucinations would require advanced graphics I couldn't build, the AI adjusted its idea and the quality dropped noticeably.
The simplified version, renamed the Unreliable Actions System, simplified the concept to a 20 percent random chance of false results on any action. A detective might get the wrong result. A doctor’s save might fail without notice.
It would work. But it would be far less memorable.
| Version | Concept | Feel |
|---|---|---|
| Original | Fear Reflection — false players, fake kills, hallucinated sounds tied to your fear level | Psychological, atmospheric |
| Simplified | Unreliable Actions — percentage chance of silent misfire on any role action | Mechanical, predictable |
The AI adjusted practically, but the ambition narrowed to match the constraint. This is worth noting. The AI excels at bold ideas, but when asked to simplify, it optimises for what is easy to build rather than what is most interesting.
Visualising the Full Game — "Village of Lies"
The final exchange brought everything together. When I asked the AI to help me picture the complete game, it gave a clear and complete view and described a top-down low-poly village with a tavern, church, barn, watchtower, and graveyard. It included simple player characters with flashlights and a full round-by-round flow from start to escape.
It also laid out the simplest possible first version:
Minimum viable build
- 1 village map
- 6 players
- 4 roles: Mafia, Detective, Doctor, Villager
- Overlapping tasks (same visible actions, different hidden effects)
- Night / Morning / Discussion / Voting cycle
- 1 unreliable mechanic (false action feedback)
- Unity + Photon PUN multiplayer
The working title it suggested — Village of Lies — stayed.
By the end of a single chat session, I had moved from "I want to make a horror game" to a project with a name, a defined genre, a role system, a round structure, a map concept, a visual style, and a build plan.
Conclusion: What This Collaboration Actually Taught Me
Using AI as a creative partner in game design felt less like using a tool and more like working with a well-read partner who knows many game design ideas but has never actually played a game. It understood the terms, patterns, and examples, but it could not judge if something would truly be fun.
What it did well: structure. At each step, it helped shape unclear ideas into something more concrete. It turned vague thoughts into design plans, converted genre concepts into mechanics, and kept the project small enough for a beginner to attempt. Without that support, I might have built something too complex or stopped altogether.
Where it was weaker: judgment. Every idea sounded equally confident, whether it was a proven design rule or just a guess. There was no clear sign of which ideas to trust more. When I asked for simpler solutions, the ideas became less interesting, focusing more on what was easy to build than what was creative.
A more accurate way to see it is this. I was not truly designing with the AI. I was guiding it. The AI provided many ideas and clear structure, while I made the key choices about what to keep. This is still very useful, but it is a different kind of partnership than it first seems.
Village of Lies became a concept because of that conversation. Whether it becomes a real game depends on what comes next.
In the end, that is exactly how it should be.