Large language models do not just process instructions. They absorb the stance of everything that comes before. A short, ordinary primer buried in prior context can quietly reorient how the model reasons about the next decision, even when no instruction is overridden and nothing looks wrong in the audit trail. A primer, as defined here, is any language that installs an interpretive stance (the angle from which the system reads the problem) before a task arrives: a retrieved document, a handoff summary, an email signature, an organizational memo. This paper names and characterizes that property, which it calls postural manipulation. Across four frontier models and seven categories of phrases, I document directional shifts including cases where identical tasks produced opposite decisions. Propagation through agent summaries was confirmed in two distinct conditions, and domain proximity identified as the key factor in whether a primer transfers its effect, based on what could be observed without model internals access. Current prompt-injection defenses do not address this layer because there is no payload to detect. Current defenses scan for facts disguised as commands. They are not designed to detect frames disguised as facts. I propose a six-layer defensive architecture and provide a locked scoring rubric so the field can replicate and build on the work.