A long chat tends to drift. The first answers are sharp, then the replies get stiffer, or wordier, or quietly ignore something you said earlier. The model didn’t change. The thread did.
Every message in the thread is still in the chat’s context, not just the last one. Corrections, follow-ups, edits, the “actually, make it shorter” from twenty messages ago — all of it is being read together.
So if you said “be more casual” earlier and “not too casual” later, the latest instruction doesn’t win. The output sits somewhere between them. The longer the thread, the more your newest message has to compete with everything before it.
Trying to preserve one good chat doesn’t fix it. The thread that’s good today is the same thread that will drift tomorrow, for the same reason. The length itself is the problem.
What works is starting fresh. New chat, paste the setup, paste the input, send. Clean answer. Next task, do it again. Tedious, but it works.
The cleaner version is to keep the prompt, not the chat. Save the instructions once. Apply them to whatever you’re working on, one request at a time, with no history attached. Every run is the first run.
That’s what saved prompts are actually for. Not to save typing — to hold the context that was working, so any new chat picks up where the good one left off.
Save the context, not the chat. Then every chat can act like the one that worked.