This week I experimented with building an AI game designer to help flesh out project ideas. I transferred a bunch of handwritten notes from different books into digital form to feed them as instructions to LLMs, and then tested several models.

The result of generating ideas from scratch or from a given theme turned out unsatisfying. For some reason, almost every AI tends to blow projects up to huge scales, almost always adds time-manipulation mechanics or unrealistic systems, and generally multiplies the scope of work.

For a test with refining an existing concept, I picked an old idea about a magic school. Something like a mini-mobile version of "Hogwarts", but without advanced combat, storylines, etc.-just short runs into 2–3 generated dungeons plus a hub in the form of the school, with students and teachers acting as quest-givers.

Overall, it does point out flaws in the GDD and can suggest some interesting improvements (especially when guided by design best practices), but its biggest weakness remains the tendency to endlessly “improve” the concept and its fear of minimalism/simplicity.

In the end, based on GPT-4.1 and GPT-5, it’s possible to do some GDD refinement and detail generation for existing concepts, especially when following clear game design rules.

Will I make this mini-Hogwarts? For now I’m practicing with planning - I’d like to learn to structure game projects from start to finish at least 80%. But if I manage to create a full plan, I’ll release at least a test alpha build in the store to see if there’s an audience for the idea.

p.s.: Notes are based on "Jesse Schell - The Art of Game Design".
p.p.s.: Screenshot from the sandbox game "Tiny Glade", used as inspiration for designing the school’s levels.