Instruction Hierarchy

A language model receives instructions from several directions at once: platform rules, organizational policy, the application's prompt, the user's request, and any text inside retrieved documents. Without an explicit hierarchy these compete on position and phrasing rather than on authority, which is what makes prompt injection viable — an instruction planted in a document can read as more recent and more specific than the policy it contradicts. An instruction hierarchy resolves the conflict in advance by assigning tiers, so a rule about redaction or scope cannot be overridden by anything arriving later regardless of how it is worded.

The Skill Manager loads skills in five fixed tiers, with governance first and style last. Governance skills are force-loadable and cannot be overridden by anything that loads after them. Because skills are pre-indexed rather than assembled at runtime, the hierarchy is a property of the context window before the first record is retrieved.