what actually goes in the system prompt vs just typing it in the message
still fuzzy on this one. when people say put it in the system prompt vs just the user message, whats the actual difference in practice? does the model weight the system prompt more or is it just organization for us. asking bc my agent keeps ignoring instructions i shove into the user message halfway through a session
system prompt is the rules that hold every message, persona and constraints. the message is the actual task. if you put the task in system prompt it sometimes ignore it after few turns in my testing. small difference for a chat but for a agent it matter a lot
this matches my testing too. one thing i would add, if the task drifts out of system prompt after few turns it is often because the model summarize its own history and the persona get compressed away. we pin the constraints again in a short reminder every ~10 turns and it hold much better. is your case chat or a long agent loop)
off topic but bpebert if you crack the write-like-me thing lmk, ive been fighting the press release voice for months