The new world of prototyping with AI

Coding with AI is definitely a lot of fun, but it sure doesn't require less knowledge. There are countless videos on YouTube showing people "vibe coding" or whatever, but I can guarantee that what they have likely doesn't scale very well, or is perhaps insecure.

ADHD waiting aside, because I was just in the midst of a six-minute wait for some changes I asked for, Claude can do some weird things. For example, I have some backend code that examines the security token and figures out if it's a valid user making the request. It decided to just stub out the user (with userID = 1) instead of actually making sure that it existed. People keep saying that it will get better, but will it? If you look at code out in the world, most of it isn't very good, and the stuff in public is what the AI trains on. I guess we'll see.

I've been working on a prototype for... something that needs a better name than "social network," because I feel like this is something different. It's for me, first and foremost. It's photo journaling and microblogging, which is what socials used to be before they became algorithm driven. It's obviously not something that can exist for free, because cloud stuff isn't free, but what if you could as a paying user invite others to use it in a read-only manner for free? I would pay for that, but who knows if others would.

Hilariously, and representative of my lack of follow-through, I made the first commit to this code more than five years ago. I didn't really revisit it until last summer, briefly, before picking it up again after getting a Claud subscription. I'm tired of messing with the forums, and needed something different. But importantly, I'm doing it for me, with possible (if unlikely) commercial potential. That's what I've always done with the forums, for 25+ years, and I imagine that's why they still exist.

The prototyping, now that I've started it in earnest, goes pretty fast. I'm still trying to find the level at which I'm willing to back off and let the AI leave it in a "good enough" spot. It's not easy, but Claude does a nice job summarizing what it changed and what approach it used. And if I think it sucks, provided I'm committing frequently enough, I can just reverse what it does.  For example, that six minutes that I had to wait involved changing six files, adding about 100 lines and deleting 30.

I've got some time off coming soon, and I hope that I can just bury myself in this now and then.