#system prompting

1 messages ยท Page 1 of 1 (latest)

slate fulcrum
#

For sure, the ideal state is no system prompt at all. It's only for the models that seem to really need it like Gemini, and in the context of MCP it seems less likely for it to be needed, since in that situation there's an established set of tools that you've given it.

Where Gemini struggles is "you have no specialized tools, but I have given you tools that give you tools, figure it out" - OpenAI and Anthropic deal with that just fine, but Gemini needs to be given a lot more info.

It's also possible that for MCP we add a tool like "readme" that gives it directions, but again I don't know if we'll need it at all

#

(Another reason "no system prompt" is preferable is to minimize token cost, which is something I'm heavily weighing in the tool calling scheme itself)

lean crypt
#

My best guess is a _readme builtin tool, possibly mandatory ๐Ÿ™‚

#

_readme_if_you_are_gemini ๐Ÿ˜›

slate fulcrum
#

dagger_for_dummies