Chat-first LLM tools make staying in one long conversation easier than starting a new one. So people stay, the thread fills up with unrelated work, and the model quietly gets worse at helping.
The fix I'd make: notice when the topic changes and offer a one-click split that carries the context over. Starting fresh should cost less than staying.
Every reply in a thread is a snapshot: the model, everything in the context so far, and your latest message, all at once. So the thread is really a record of what the model was working with each time it answered. When the thread...