Counter-Attractor Prompting

Language models generate autoregressively, so what is already in the context pulls what comes next toward it. Prohibitions (“do not invent quotes”) put the failure into the context and pull toward it. Counter-Attractor Prompting starts by naming the task’s known failure modes and then fills the context with counter-examples and a structured self-audit, so the pull runs the other way. The protocol is evaluated with controlled experiments against the plain and prohibition-style prompts for the same task. A write-up with the experiments is in preparation.