AI Periwinkle
Inspiration
AI Periwinkle began immediately after I finished my first game, The Hardest Game Ever Created. That was the first game I had ever made, and I completed it in less than 24 hours despite having no traditional game development experience.
That experience made me ask a bigger question:
Could GPT-5.6 and Codex create an entire game for me while keeping the puzzles hidden from me?
I did not want to design every room or know every solution in advance. I wanted to experience the project like a video game tester. Testing games always sounded like a dream job when I was a kid, and this experiment gave me a modern version of that experience.
My creative direction was intentionally simple. I asked for a calm, retro, first-person puzzle game with ten levels, no combat, no time limits, and an unhinged AI clown operating the facility. The final reveal would be that the clown was actually the player’s father.
Everything else was left to GPT-5.6 and Codex.
[ \text{My idea} + \text{GPT-5.6} + \text{Codex} + \text{blind playtesting} = \text{a finished game} ]
How I Built It
GPT-5.6 acted as the creative director and project architect. It developed the structure, established the technical requirements, protected the puzzle solutions from me, and wrote detailed development prompts for Codex.
Codex acted as the game studio.
It created the Godot project, wrote the GDScript, built the interaction systems, designed the rooms, implemented the puzzles, created the dialogue system, added saving and resetting, tested the logic, and prepared each level for me to play.
The game was developed one level at a time.
After Codex finished a level, it returned control of the computer to me with Godot open. I pressed Play and entered the room without knowing the puzzle or its solution. I had to complete each level before development could continue.
The workflow was:
- GPT-5.6 planned the next development stage.
- Codex created and tested the level.
- I played it as a blind tester.
- I reported the problems I personally encountered.
- Codex repaired those problems before moving forward.
Codex completed the core development of all ten levels in approximately two hours of active coding time. My playtesting took additional time because I had to solve every puzzle myself.
The project was created in Godot on a Microsoft Surface 7. The hardware was not especially powerful, but Codex handled the technical work that would normally require extensive programming knowledge.
Challenges
One unexpected challenge was my keyboard. The top letter row was broken, which meant I could not use standard controls such as W, E, or R.
Codex adapted the entire game to use:
- D to move forward
- X to move backward
- Z and C to move sideways
- F to interact
- V to reset a room
We also encountered several real development problems during testing.
Some geometric symbols became corrupted because of text encoding and unsupported font characters. Codex replaced the fragile symbols with reliable rendered shapes.
Several puzzle instruction panels were initially too large, clipped, or blocked by physical controls. These problems were difficult for automated testing to detect because the underlying puzzle logic still worked. My role as the human tester was essential. I could immediately see that information was technically present but visually unreadable.
The dialogue system also allowed three lines of text to overflow a box designed for only two lines. Codex repaired the shared system so longer messages automatically continued across multiple pages.
These challenges taught me that functioning code is not the same as a good player experience. Automated validation can confirm that a puzzle has a solution, but a person still needs to determine whether the clues are readable, whether the solution feels fair, and whether the game is enjoyable.
What I Learned
The biggest lesson was that I did not need to understand the complete architecture before starting.
I could describe the experience I wanted, test the results, and communicate what felt right or wrong. GPT-5.6 and Codex translated that feedback into functioning software.
I also learned that AI development does not remove the human role. It changes it.
I was not manually writing every script, but I was still making the game. I chose the premise, established the tone, tested every level, identified problems, approved the puzzles, and decided when the experience felt complete.
Most importantly, I learned that coding is becoming accessible to people who were previously excluded by the technical barrier.
AI Periwinkle was built by a first-time game developer, on an ordinary laptop, through conversation and playtesting.
That is what Codex represents to me:
Coding is no longer reserved for programmers. If you can imagine something, explain it, test it, and improve it, you can build it.
Log in or sign up for Devpost to join the conversation.