Does anyone have a reliable strategy

[raw saddle]

will see what i can do

ideally you want some global instruction that is added to system prompt in text mode, one in voice mode, maybe a separate one still if both, right?

[dry gust]

it might be sufficient to have a dynamic variable that tells the agent whether its in text or voice mode, i expect people do this when they use the sdk to help the agent understand the environment its operating in. so that in the prompt/knowledge i could put something like, if you are in text mode you can use markdown, emoji and follow the guidance in knowledge ABC, but if you are in voice mode you must never use emoji, markdown and you must refer to knowledge XYZ.

does the agent know the name of the file when a file is added to the knowledge, like if i refer to conversation-guidelines.md will it be able to see that filename in the context or it just has whatever is the contents of that inserted into the prompt context?

[dry gust]

i just realised i havent tried using this feature yet, does it fix the kind of issues i have like zed instead of zee, 'at' instead of '@' etc?

[raw saddle]

yep, if choosing elevenlabs likely would fix that