You're effectively steering the predictions based on adjacent vectors (and precursors from the prompt). That mental model works fine.