The point is the current model also knows nothing about the “on the job stuff”.

It’s extremely difficult(impossible?) to include every bit of relevant domain knowledge into “the prompt”