An AI tool changes a character’s jacket between scenes. I correct it, generate again, and the old jacket comes back. Now I have to guess whether it ignored me, reused an earlier reference, or lost the correction when the conversation was summarised.
That guessing is an interface problem. The tool knows which references and instructions it sent to the model. Some useful version of that information should be available to me too.
I’d like to see the active character reference, the chosen style and the constraints for this generation. If an old reference is wrong, I should be able to remove it. If a correction belongs to the whole project, I should be able to save it there rather than repeat it in each prompt.
A raw context dump would be overwhelming. A small list of named sources and decisions is enough for many cases. Give each one a scope and a way to edit it.
This also applies outside creative work. A research assistant should show which document supports a claim. A coding agent should distinguish repository instructions from text it found in an issue. When the result is wrong, I need somewhere better to start than asking the model why it did that.