I think there's a really interesting tension between what you're saying and what the parent post is saying, and it's something that's been discussed for a long time but LLMs are throwing a new wrench into it: a given user probably only uses 5-10% of the features of a piece of software, but across the whole body of users you get close to full coverage.

I use DaVinci Resolve for editing GoPro, Canon DSLR, iPhone, and Drone FPV footage. I import clips, use the razor to cut them and rearrange them, occasionally crop and zoom, and then encode them down to h264 or whatever. There are a bazillion features that I sort of know exist but haven't ever had any need to use. Could I make a tool using an LLM that covers my DaVinci workflow? Almost certainly, and I don't think it would take anywhere near 200 prompts. I've made significantly more complex bespoke data analysis pipelines for drone flight telemetry in like... a dozen prompts. But I have no illusions that what I would make as my "DaVinci Clone" would have anywhere near the functionality of DaVinci, just the tiny piece of that functionality that I use.