Currently I'm working on SNKRX's update, this is the only project I work on almost every day and have been managing to keep it going for about a month, and then I explore other projects as I see fit, like Phylactery which I ended up finishing. So as I explore potential ideas and as Opus 5.5 is more capable, I've been trusting it with running an idea mostly by itself more often. The problem is that this almost always ends up being a mistake if I let it happen at the very start of the project, it has never happened yet that I play what it generates and I go "oh, this is amazing" for more than a few moments. In these outputs there are individual things I may think are amazing and may take, but never the whole at all. If the project is further along with an initial implementation iterated by me and I let the AI run with it then, the success rate climbs but only barely. Overall it seems like letting the AI go on big runs doesn't work for me.

There are two possibilities on why this happens. First is that most ideas are bad, so exploring the possibility space of ideas will mostly lead to failure. I had Claude go over all prototypes I made in the past ~15 years (since I started making games) and it returned me that I made around 130 prototypes. I then picked 10 promising ones that weren't too far along to actually have Claude implement to a reasonable degree and played them and most of them suck. It just feels right that most ideas are bad and not particularly fun.

But the other possibility is that seeing the result with no effort on my part makes my brain reject thinking the thing is fun. A game made like this has lots of small mistakes that I wouldn't have made if I built the game myself, so as I play it, it seems like I can't ignore the mistakes and judge the game accurately. I'm kind of a perfectionist so playing something that's imperfect actually ruins the whole idea for me. So this second possibility means I'm ruining potentially good ideas by not experiencing them at their best the first time. There's usually a moment when you're iterating on a game where you play it and you go, holy shit, this is actually fun, that moment happening is important because it convinces you the idea is good, but having the AI implement a half-baked version of the idea essentially robs you of it early on. This is ultimately a compass-deadening/noisifying (#34) process.

The conclusion then is that I shouldn't do big automated runs anymore, at least not for the start of the project. The seed, the initial way the game plays, the structure of the game, it has to be iterated slowly and only then I can try to explore it in more automated ways.