I still have internal disagreements about the question of how to determine the usefulness of content on Steemit. What do curators base their decisions on?

I do, too. That's why I try to avoid the phrase "quality content". To me, the content that should be incentivized is the content that draws eyeballs - so I usually refer to "attractive content" - i.e. it attracts an audience. I know from other social media platforms "attractive content" isn't always "quality content", though it's nice when the two are aligned.

However, it's not easy to define what "attractive content" really looks like. In my own mind, one of the things I look for is "surprise", which is consistent with this:

the Thoth bot evaluates or should evaluate the informational value

I also try to look for relevance and evidence of higher order thinking: not just parroting from one source, but combining information from multiple sources and drawing conclusions from it.

You can see the prompts that ship with Thoth, here and here. I've tailored it a little bit in my own environment, but that's basically still what I'm using.

When I created the LLM integration, I included the ability for each person who runs the bot to replace those prompts if they have different ideas about what sort of content should be supported. The idea is that each person tailors it to their own preferences to build a combined, decentralized result that covers the full range of human diversity of thought.