System Prompts

The system prompt is the set of instructions that defines how the agent should think, respond, and act. It is the most important part of the agent’s configuration.

The two editor modes

In the Prompt tab, SquadOS offers two modes:

• Visualize (default): the prompt rendered as markdown, read-only.

• Edit Prompt: markdown editor with toolbar (bold, headings, lists, code, quotes) and a side Available tools panel. Clicking a tool inserts the {{tool:tool_name}} placeholder, which becomes a live reference when the tool is called. To the right of each tool there’s an info icon (ℹ️) that opens a ready-made usage suggestion — see Referencing tools in the prompt.

Next to it, the Edit with AgentMaker button opens the copilot that helps you rewrite the prompt through conversation — see AgentMaker.

What to put in the prompt

A good prompt usually has these parts:

• Role: who the agent is.
• Goal: what problem it solves.
• Context: company, product, or process information.
• Rules: what it should always do or avoid.
• Tone: communication style.
• Flow: steps it should follow.
• Tool usage: when to call tools or knowledge bases.

Simple example

You are a support assistant for Acme.
Your goal is to help customers with questions about access, billing, and account setup.

Respond clearly, concisely, and politely.
When unsure, ask before assuming.
If the question depends on internal data, use the connected knowledge base.
If the user asks for something outside support scope, explain you cannot help with that.

Best practices

• Be specific about when the agent should act.
• Prefer observable rules over vague phrases like “be nice”.
• Explain when the agent should ask questions.
• Describe limits: what it should not promise, calculate, or execute.
• Include examples when behavior is sensitive.

Referencing tools in the prompt

If the agent has tools, guide when they should be used. In the direct editor, the Available tools panel inserts the {{tool:name}} placeholder — useful to keep references consistent:

Use the {{tool:query_base}} only when the user provides the account email.
If the email is missing, ask before calling the tool.

This reduces wrong calls and improves predictability.

Ready-made usage suggestion

You don’t have to write the instruction from scratch. Click the info icon (ℹ️) next to a tool in the Available tools panel to open the Suggested prompt usage box, with two ready-to-copy texts:

• Full phrase: a natural-language instruction already written for you (e.g. “Use the search_courses tool to reach the external integration when you need the data it provides.”). Paste it into the prompt and adjust to your case.
• Tool name: just the tool’s technical name, so you can craft your own sentence.

Every tool — native, integration, or MCP — ships with its own suggestion. It’s a starting point: once pasted, tailor it to whatever makes sense for your agent.

Reviewing the prompt

After saving, test the agent with real questions. Adjust the prompt whenever you notice:

• responses too generic;
• tool calls without enough data;
• inappropriate tone;
• reluctance to ask when info is missing;
• responses out of scope.