The OpenAI instruction hierarchy prioritizes the **Developer Message** (often referred to as the system message) over the **User Message**.
**How it works:** 1. **Priority:** The Developer Message sets the foundational context, rules, and persona for the AI. It has higher authority than subsequent User Messages. 2. **Function for AI Agents:** For AI agents (such as those using the Model Context Protocol or similar tool-use frameworks), the Developer Message typically defines: * **Tool Definitions:** Which tools are available and how to use them. * **Routing Logic:** How to decompose questions or route them to specific sub-questions or services. * **Behavioral Constraints:** Rules for grounding responses in real tools, handling errors, or formatting outputs. 3. **Execution:** When a User Message is received, the AI evaluates it against the constraints and instructions set by the Developer Message to determine the appropriate action (e.g., calling a specific tool, generating a response, or decomposing the query further).
In the context of your project, **Liminality**, the Developer Message would likely contain the logic for decomposing questions, grounding them to tools like Acme CRM or Motherboard, and defining the routing rules for the MCP server.
This answer was worked out by Liminality — Physea's engine that decomposes a request, grounds each part to a real tool, and returns a reusable, checkable route.
Connect it over MCP: https://mcp.physea.ai/mcp · physea.ai