This is something that is very hard to do properly, as system prompts control not only instructions but also the expected output format.
If we allow users to customize now, we are signing up for a constant barrage of frustrated users because our updates will break their site whenever we need to update the expected output format that is parsed for features like summaries, helpers, etc.
One way around this is to move everything toward strict formats, using Guidance / functional calls / structured outputs. However, the ecosystem is only now reaching the maturity level where that is available in most providers and on self-hosted options. This is important because Discourse AI does not lock communities to OpenAI; it allows people to pick and choose from thousands of compatible LLMs.
All this to say that it is something we want to do, but we can’t half-ass the implementation. That means we need the entire AI ecosystem to catch up, and that is only happening now.