Yeah, I've been thinking a bit about the multi shot thing. I've had great results from Claude by "faking" the previous conversation to include example question/answer pairs.

With LLM you can do that using the undocumented Python Conversation API, but it's undocumented for a reason (I don't think it's good enough yet). You could also fake a previous conversation through the CLI tool but that is VERY undocumented - you would have to write fake rows into the SQLite database!

I also want to support the Claude thing where you can prefill the start of the response - amazing for things like forcing HTML by refilling an HTML doctype.