It would be helpful to have finer grained control over the system_message. Right now, you're stuck with append or replace only.

The current append mode gives the model access to tools and skills, but also adds a lot of opinionated constraints that don't make sense for general purpose agents. Instructions such as "keep responses to 3 sentences or less", and "ALWAYS use the fetch_documentation tool", can mean that asking simple questions such as "What skills do you have?" can be very unreliable.

Conversely, the current replace mode gives the model access to tools and skills, and very little else. However, some of the append mode prompt can be genuinely useful; just not all of it all of the time.

Ideally, I'd love something that lets me configure which parts of the prompt that I want, preferably at a semantic level. So, if I wanted the model to give lengthier explanations, or to set a different tone, or to output markdown, or to stop outputting emojis, then I could do so without having to recreate the entire system message.

I understand that this could be tricky to do as a one-size-fits-all, but perhaps a prompt builder with sensible defaults, or model-specific prompt templates would be the way to go.