Repeated agreement, repeated violation
If the model keeps violating the same rule after correction, you are in a re-prompt spiral. The classic shape: human corrects, model agrees, model violates again, human corrects more emphatically, model agrees more enthusiastically, model violates again. Each cycle adds tokens but does not change behavior.
Why it happens
Old context is still in the window, still influencing attention. Adding more reminders amplifies noise without removing the original misdirection. The session has lost the attention battle, and more prompting is the wrong tool.
Break the spiral
Stop adding reminders. Capture the live truth in a durable note. Start a fresh session. Reload the note and current source. Continue from a clean prefix. The fresh session usually behaves correctly within one turn — proof that the problem was context pollution, not capability.