Most posts on this blog are about the category. This one is about us, specifically, and what building in this space has actually looked like from the inside.
The playtest loop took longer than the generation did
We assumed the hard part would be getting a model to produce working code and art. It wasn't. The hard part was building the automated loop that plays a generated build, notices what's actually wrong, not just what crashed, and feeds that back in a way the next generation pass can act on. That infrastructure, invisible in every demo, ended up being most of the engineering time.
We underestimated how much people want to just watch it work
The playable demo on our home page, watching the machine visibly build a game from a prompt, gets more engagement than any written explanation of the pipeline. People don't want to read that it plays its own build and fixes bugs, they want to see the log say "playtested, no errors" in front of them. We built more of that visible, honest theater than we originally planned, because it's what actually earns trust.
Pricing honestly is harder than pricing cleverly
It would be easier to price this like unlimited-everything and figure out the unit economics later, the way a lot of funded startups do with a runway to burn. We don't have that runway, so we had to be honest about metering generation from day one, which is a less exciting pitch but a business that can actually survive its own success instead of being strangled by it.
We're still not sure exactly what a "generation credit" should cost
The founding member pricing on this site is deliberately structured around results, enough credits to ship a first game, guaranteed, rather than a specific number, because we don't yet know the exact cost-per-generation well enough to publish a confident credit table. We'd rather say that plainly than publish a number we'd have to quietly walk back later.
What we'd tell ourselves a year ago
Ship the honest, smaller version sooner. We spent time polishing things that mattered less than we thought, and not enough time on the boring infrastructure, playtesting, error recovery, that actually determines whether the product works when nobody's watching. If you're building something similar, that trade favors boring infrastructure every time.