Attach several pictures and the assistant now knows which is which: each one is labelled in the turn, so "change the second one" works, and "do each one" runs across all of them. Images the answer refers to appear inline where they are mentioned, and the ones it did not place are shown with the reply instead of being dropped.