Early idea to testable “V0”
Over the past few weeks, I’ve been testing how AI can support small creative studios through early-stage game development. The goal is to see whether AI can make the path from an early idea to a clear, testable version – what we call a “V0” – more structured, and less risky.
The project is exciting and complex, and has already begun to reveal an increasing tension: the better AI becomes at structuring creative work, the more careful we need to be that it doesn’t start pre-imagining the creative decisions themselves. What do I mean by pre-imagining? Giving us a false sense of certainty that creative decisions can be made before they’ve been tested against the most complicated system of all, humans. Particularly creative humans and, in our case, playtesters.
AI is very good at turning vague ideas into something buildable
Yes, AI can certainly help creators move from a loose concept to a structured game design document (GDD). We released a “Game Design Skill” tailored to Decentraland, the virtual world I work in, so that studios can use their prefered AI agent as a conversational design partner, then as an autonomous prototyper. The Skill leads them through questions to define the core loop, mechanics, player goals and social dynamics. Then defines and builds early tests. It’s particularly useful at identifying inconsistencies, like asking about missing ideas about game progression, and maintaining context across a long design process.
The resulting GDDs and tests have generally been clear enough to move directly into development. I can attach the output of a gdd.md to a purchase order and commission the V0 directly – all suggesting that AI can remove a meaningful amount of creative admin and back and forth from small teams. To draw out the level of detail I’d need for a commision, there would previously be a long email trail of “how will this work”, “what’s your retention strategy”, “what will your player do first”. The GDD Skill cuts through all of this to a complete, easily reviewable plan and test scene.
Structure is valuable, but it can easily lose its soul
The biggest risk we have seen is almost the inverse of that strength. When an AI system is very good at asking the “right” design questions and proposing plausible answers, it can make the creative process feel more linear and prediactable than it really is.
A studio can quickly move through a sequence of questions and arrive at a coherent design, but coherence is not the same thing as originality, nor something that will actually work in practice. There is a danger that creators start selecting from suggestions rather than discovering the game themselves.
One tester described the process as significantly faster, but felt that the resulting idea contained less of their own creative “soul.”
So I, with my excellent team of developers, need to get better at creating Skills that feel like an inspiring creative conversation, not like completing a form with intelligent autocomplete. It should create more space for experimentation, and wild unproven ideas, not less.
The AI needs to explicitly encourage disagreement
For AI workflows to push us to the original and the human, we need to build in the space and permission for studios to say:
“This mechanic just doesn’t feel right.”
“I want to explore a completely different direction.”
“I know what you’re telling me and I don’t know how to explain why, but I need you instead to do this. I know the data doesn’t back it up. I want to do it anyway.”
Throughout the history of creativity, good ideas appear as tangents, contradictions or even dreams (I still remember my chemistry lesson on Kekulé’s the ring structure of benzene). An AI system designed around efficiently reaching a finished document can very thoroughly optimise those moments away.
The workflow around the AI matters as much as the AI itself
Something else that has become very obvious through testing is how much of the experience has almost nothing to do with the quality of the model.
We have had moments where the AI itself was doing a perfectly good job, but the surrounding workflow was making the whole thing feel harder than it needed to. A human coming back to an existing game might naturally be thinking about the next version, while the assistant assumes they are starting again. The agent can store files in unfathomable places, or the workflow can tell someone to create a project when they already have one.
None of these are particularly exciting problems, but they matter enormously for creatives to feel trust, and not frustration, in their workflow.
There’s a temptation with AI tools to focus on whether the assistant gave the “right” answer. From a studio’s point of view, that is only part of the experience. They care whether it understood where they were in the process, whether it remembered what they had already done. And importantly, did it help them get to the next stage in a project without making them jump through seemingly random or excessive hoops.
The more capable the AI becomes, the more noticeable these workflow failures are. You start to expect the whole experience to understand you, not just the conversation window.
We probably care more about attention than tokens
We also went into this expecting token usage to be a fairly significant concern. So far, it has not emerged as the thing creators care most about. The bigger cost we’re seeing is attention.
If the assistant gives you a wall of text, asks five questions at once, explains something you already understand and then repeats the decisions you made ten minutes ago, the process starts to get annoying very quickly. Even if the actual financial cost of that exchange is tiny, the creative cost is not. That feels particularly important when you are working with studios. The point of the tool is to help them spend more of their energy on the game, not on managing the AI.
So there’s a useful distinction here between making an interaction shorter and making it more efficient. A good design conversation might still be long. We might want people to spend time exploring an idea. But each part of the conversation should feel like it is earning its place, while inspiring a studio to keep creating.
What I think we are actually building is creative scaffolding
The thing I keep coming back to is that the most useful role for AI here is probably not “game designer.”
It is much closer to scaffolding around the creative process.
There is a lot of work involved in getting from an idea to a game: remembering constraints; breaking things into discrete, concrete steps; finding contradictions; making sure an important design question has not quietly disappeared; working out what should be tested first. AI is good at carrying a lot of that weight.
If the system is holding the structure, the creator has more space to think about the parts that actually require what some people call “taste”. I’m not a fan of that word, and we’ll probably unpack that in another article. But let’s take “taste” as meaning human questions like: is this fun? Is this surprising? Is this something I care about making? Is there another version of this idea that nobody has tried?
The main danger is that this scaffolding becomes the actual building. If the fastest route through the process is simply to accept every suggestion the assistant makes, then we can very efficiently produce something coherent that nobody feels particularly attached to. And that no one will find fun or inspiring or moving to play. That’s the opposite of a successful creative workflow.
That feels like the big, official learning for me so far. The value of AI in a creative workflow is not that it can make more of the decisions. Instead, it should remove enough friction around those decisions that creators have more time and space to make the interesting ones themselves.
