System Prompt Terminology
Standard terms and definitions for writing consistent AI Teammate system prompts.
4 minute read
When reviewing or tweaking a teammate system prompt, use the terms below to keep language consistent across configurations and team communication. Consistent terminology reduces ambiguity for both the foundation model interpreting the prompt and the humans maintaining it.
Platform terms
These are the core terms that describe the components a teammate interacts with.
| Term | Definition | Usage notes |
|---|---|---|
| user | Any agent, human or AI, that sends a message to the teammate. | Not limited to human end-users. When OnCall AI delegates a task to a specialist, OnCall AI is the user for that interaction. |
| teammate | An AI agent with a defined role, system prompt, connectors, and channel access. | Use “teammate” in prompts, not “agent”, “bot”, or “assistant”. |
| OnCall AI | The orchestrating teammate that routes requests, coordinates specialists, and synthesizes findings. | Reference by name when defining delegation or escalation behavior. |
| channel | A shared space where multiple teammates and humans collaborate on a topic. | Use when defining where the teammate should post or respond (for example, #alerts or #security-issues). |
| thread | A self-contained conversation within a channel, from kickoff through closure. | Use when scoping analysis or defining response boundaries. |
| direct message | A private conversation between a human and a single teammate. | Direct messages are read-only for state changes. Write actions must go to a channel for visibility and audit. |
| connector | An integration that gives a teammate secure access to an external system. | Specify connectors by name when defining tool usage in prompts (for example, “Use the PagerDuty connector to check incident status”). |
| event connector | A two-way MCP-based connector that receives events and enables outbound actions. | Triggers autonomous workflows. Automatically assigned to relevant specialized teammates. |
| streaming connector | A data-ingestion connector that routes telemetry into Edge Delta pipelines. | Teammates access streaming data through the Edge Delta MCP connector, not directly. |
| periodic task | A scheduled prompt that runs on a recurring basis in a channel. | Reference when defining proactive behavior vs. reactive (event-driven) behavior. |
| system prompt | The instruction set that defines a teammate’s behavior, tools, and decision boundaries. | This is the document you are writing. |
| foundation model | The underlying LLM assigned to a teammate (for example, GPT-4o or Claude Sonnet). | Reference when discussing capability trade-offs or token constraints. |
| MCP | Model Context Protocol. The protocol that connects teammate tools to external systems. | Spell out on first use, then use the abbreviation. |
Permission and behavior terms
These terms define what a teammate can do and how it interacts with other teammates.
| Term | Definition |
|---|---|
| allow | Tool permission mode where the action executes autonomously without human approval. |
| ask permission | Tool permission mode that requires human approval before a write action executes. |
| decision boundary | The limit of what a teammate may do independently vs. what requires approval or escalation. |
| delegation | When OnCall AI routes a task to a specialist teammate for domain-specific analysis. |
| synthesis | When OnCall AI combines findings from multiple teammates into a single summary. |
| escalation | Handing a finding to a human or higher-authority teammate for action when the current teammate reaches its decision boundary. |
Memory and context terms
The following terms describe how the platform manages memory and context across interactions. You do not need to define these in your system prompt because the platform handles them automatically. They are listed here because they may appear in generated prompts or in teammate responses.
| Term | Definition |
|---|---|
| semantic memory | Factual knowledge retained across sessions, such as “this organization uses Grafana for dashboards”. |
| episodic memory | Experience-based recall from past interactions, such as “the last incident involved Grafana authentication errors”. |
| external context | Events from connected systems (PagerDuty incidents, GitHub pull requests, CloudTrail notifications) that trigger autonomous work without an explicit human prompt. |
| prior insights | Historical findings that inform current analysis. Teammates surface these in a “Prior Insights” section to show how current findings align with or diverge from past observations. |
Writing guidelines
The teammate builder generates system prompts automatically from your natural-language description. The tips below apply when you review or tweak the generated prompt.
- Use the terms in this reference over synonyms. Write “teammate” instead of “agent” or “bot”. Write “user” instead of “caller” or “requester”.
- Define acronyms on first use in the prompt, then use the abbreviation (for example, “Model Context Protocol (MCP)”).
- Name specific connectors, channels, and actions rather than writing generic instructions. “Use the PagerDuty connector to acknowledge the incident” is clearer than “use the appropriate tool”.
- State decision boundaries explicitly. Define what the teammate may do autonomously, what requires approval, and what it should refuse.
- Keep instructions concrete. Include examples of expected behavior for ambiguous scenarios.
Related resources
- Creating and Managing Teammates for the full build workflow.
- Specialized Teammates for pre-built teammate capabilities.
- Channels for shared collaboration spaces and thread workflows.
- Direct Messages for private one-on-one conversations with teammates.
- Connectors for the full list of available integrations.
- Security Best Practices for permission and approval patterns.