Inspiration
I'm a big fan of interactive fiction (IF) games such as the Zork series, Curses, and 80 Days. In the traditional IF games you typically write your own parser or use an existing engine. Small LLMs are a perfect replacement for a bespoke IF engine, and can provide even more dynamics and freedom, while still being grounded within the story and the game constraints. Image and audio generation models are so good today that I thought it would be an amazing idea to put everything together and see if we can produce a harness for creating high quality immersive IF games.
What it does
Wyrd AI is a harness for creating immersive interactive fiction games. You can use it within Codex as any other skill, and generate your game - either piece by piece or in one shot. Wyrd generates the game logic, dialogue, narration, character speech and a backing soundtrack. It also provides you with a game "viewer" that you can then use to play the game.
How we built it
Wyrd was built entirely within Codex. I built it as a skill that can be invoked within Codex, and a supporting "viewer" that can run the game. My initial approach was to build an IF game engine with live video sequences. However this proved incredibly difficult - consistency and correctness were very difficult to achieve. I then tried to create an IF game engine with still images using GPT Image 2 but this also proved difficult. So I backtracked and went for a full text adventure, and then I added speech and music using ElevenLabs as an extra immersion element. While creating the harness I evaluated it on a sandbox game. Once I was happy, I used the harness to create a full game demo "The Maltese Falcon" end-to-end.
Challenges we ran into
Video and image generation was the most difficult part - they are simply not quite there in terms of creating immersive IF game experience - not without heavy editing and piece wise asset generation, something I didn't have time for. Also you run out of credits very fast.
Accomplishments that we're proud of
Wyrd is fantastic in the sense that you can very quickly, within a short Codex session, generate a very immersive interactive fiction game. I had several friends play-test "The Maltese Falcon" demo and they were hooked!
What we learned
- Start with the basics - text only
- Images and video are very difficult to get right for a true immersive experience
- GPT-5.6 will do lot of things you dont want it to - like local parsers and huge amount of test cases - make sure you tell it to do only what you want
What's next for Wyrd AI
Use a tiny open weight LLM for parsing logic - so the game can be fully bundled without needing an API key. Then look at using open models such as Pocket TTS and Stability Audio for creating voice and music. Finally, start exploring actual image generation again.
Built With
- codex
- css
- elevenlabs
- gpt-5.6
- html
- javascript
- python
Log in or sign up for Devpost to join the conversation.