When a chat with an AI keeps drifting off topic, do not rush to start over: in the same conversation, pin down the goal, the conclusions already agreed, and the requirements that must not change. Most drift pulls back. The cause is rarely a suddenly dumber model; in a long conversation the early instructions get diluted by later chatter and second thoughts, and the model ends up following the most recent few turns.
Why long chats drift
The longer a conversation runs, the more the model has to juggle, and its attention is not evenly spread: the latest messages weigh the most, while the goal, format and red lines set at the start can be buried after a few dozen turns of new topics. Add the common habit of agreeing and adapting, and it treats a passing new idea of yours as a new direction, even serving back proposals you had explicitly rejected. The sneakiest drift is gradual: each reply is only slightly off, every single step looks connected, and only at the end do you notice the whole task has quietly changed.
Three steps to pull it back
Step one: make it restate the state of play. Ask directly: what was the original problem, which conclusions are settled, and which requirements are fixed. Wherever its summary is wrong is exactly where the drift started, and correcting that beats vaguely pushing on. Step two: turn the anchors into a short fixed block of text. Condense the goal, the settled conclusions and the hard requirements into a few lines, then re-paste that block every so often, or quote it before each new request, realigning the direction on a schedule. Step three: know when to cut. If the conversation is so long it cannot even summarise itself cleanly, or the direction has already changed twice, do not rescue it by force. Ask it to produce a summary of the conclusions that are still valid, carry that summary into a fresh conversation, and state plainly which points are settled and not to be reopened. A new chat plus a clean summary is usually faster than endless corrections in the old one.
When not pulling it back is the right call
Sometimes the task itself genuinely changed mid-chat, from drafting a plan to building a budget, for instance. That is not drift but a new topic, and anchoring to the old goal only gets in the way; start fresh. And if the model keeps overturning the same settled conclusion after two corrections, the context is probably too polluted to save, and migrating with a summary saves more time than arguing. One habit covers most cases: in any long, important conversation, ask for a summary halfway through. It doubles as a record and as a fresh anchor for everything that follows.
