I built this after repeatedly running into the same issue in longer task-oriented conversations: the model gradually drifts into explanation mode and becomes verbose, speculative, and initiative-heavy.
The idea is to initialize the conversation in a stable execution mode so responses stay structured and operational over longer threads.
The README shows the same prompt in two conversations:
– default model behavior
– the same prompt with the snapshot applied
Curious if others here have run into the same drift in longer conversations.
