Policy-as-Prompt: Turning AI Governance Rules into Guardrails for AI Agents