Drift can be a sequence of reasonable-looking changes
A useful working definition of drift is a growing departure from the purpose, facts or constraints that should still govern the work. The model might answer a narrower question than you asked, revive a rejected option or turn a tentative idea into an accepted decision. No single invented fact is required.
That is why a conversation can feel productive while the project loses its direction. The last answer fits the last exchange. What matters is whether it still fits the job.
An early question behind Good Remedy was how a workflow made from prompt steps would hold together under use. A well-written opening instruction is a beginning. The business will change, corrections will accumulate and somebody else will eventually take over.
Why it happens
The model acts from the information and instructions available to it at that point. A long history may contain old decisions, new decisions and unresolved suggestions without making their different status clear. A shortened handoff may preserve the topic but lose the reason a choice was rejected.
Anthropic describes context as a limited resource and reports declining retrieval precision in long contexts. Its engineering guidance addresses how to select and maintain useful information across an agent's work. The practical implication is that putting a fact somewhere in a long conversation does not ensure it will govern every later answer.
Other causes sit outside the model. The wrong file may have been retrieved. Two systems may disagree about what is current. The person giving the next instruction may have changed the goal without realizing what depended on it. Investigate the actual departure before choosing a repair.
What catches it before the next handoff
Compare the proposed result with a short, current description of the work: the outcome being pursued, the decisions already made and the constraints that still apply. Keep the underlying evidence available so that the summary can be checked.
When something changes, identify what it affects. A new delivery date may belong in the plan, the proposal and the customer update. An instruction to remember the change is less useful than seeing those affected items brought up to date.
Try a handoff to someone who was not in the conversation. Can they identify the current version, understand what remains open and take the next step? Also test one ordinary change. A system that handles the original request but loses a correction has not yet shown continuity.
Keep the work moving
A disclaimer saying that AI can make mistakes leaves the coordination problem in your lap. The useful response is to make departures easier to see, corrections easier to apply and continuation easier for the next person.
The amount of checking should fit the consequence. A brainstorming note and a customer commitment do not need identical handling. Preserve the useful work, isolate the part that changed and continue from the corrected state.
The question to ask is whether the project is becoming more dependable as it develops. If every additional conversation makes it harder to know what is true now, the operation needs attention as well as the prompt.
Put the idea to work
If your workflow automation keeps losing corrections, start with one failed handoff. The guides examine how AI implementation and system integration can preserve the current work across teams.