Originally published as an Advisory Hour Substack post on 2026-09-09. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
This is an early build of a game tentatively titled "Fossil Interface”. I’ve not tried building a metroid style platformer yet, and so this was a step in that direction.
What was unexpected is that even this kind of game still follows the same build inefficiencies that I saw with the other games. There’s something to be said about doing the same sort of thing several dozen, if not hundreds of times. You begin to see obscure patterns in the approaches.
I had ChatGPT pull together the clone’s inspired characters and NPCs I hoped to introduce into the game. This highlights some of the build ideas and such. You can drop this image into any AI tool and build out your own variations and sprite concepts from this map.
The concept art for the main character and the various assortment of gadgetry that I have in mind.
There’s still a long road ahead. Further, I see several interesting problems in my process that can be overcome. Lot of times I’m trying to build up the game world and the game and each character all at the same time. This is just too much for today’s AI models. A little methodical breakdown does seem to help substantially. I’ve written before about the power of dedicated character and level labs. Here’s an example of one of those labs for just one of the rooms in the project. It allows for specific conversation with more accurate feedback per level. What I find so interesting about this is that these labs were created on the fly during the build out process. Each gamelab project is a unique game engine unto itself.
In the case of this image, the feedback I’d provide to the model would be around solving the solid paint bucket fills of the interior cavities of the sprites, and also the lighting of the scene. The project is getting close to playable. It’s not there quite yet.
This clip below is from a little later in the build. You can see how these things improve, evolve and change over the iterations.
And in some areas of the lab, you can see where a scene directionally starts to achieve it’s form. In many ways it reminds me of the messy work of working with lumber. You have piles of wood and tools. Then out of the chaos of the saws, sanding, hammers and wood screws you begin to see the shape of the object you held only in your mind’s eye.
Model Advance
In the midst of working on this project with 5.6sol, Astra from OpenAI was released. I had Astra take a crack at solving for this to see how it would do. It’s made me realize that the models as they continue to increase in intelligence bring into reach even more incredible ideas.
I don’t see myself completing this project now until the cost per token of an Astra class model drops significantly. Here is some concept art from the project.
What I did realize is this art style could make for an interesting video episode putting some of the sweet “local ai” to work. The result is great! I’ve included it below. I expect I’ll explore this direction more.
It’s time to try reaching for a minute plus long video.
Connected work
- GameLab: Browse the playable experiments and build records around Fossil Interface.
- Mario Walker: See another platformer that grew faster than one person could test it.
- Mario Walker Character Lab: Inspect a dedicated character workspace built for focused visual review.