跳到主要内容

System Prompt

The system prompt is the standing instruction you give the agent. The model reads it before every reply. It is the single biggest lever you have over how an agent behaves.

Where to write it​

In the agent editor → Basic Info section → the System Prompt field. It is a multiline text field embedded within the Basic Info card, below the model selectors.

Placeholder text: "What should this agent do? Include tone and any must-do steps."

In Advanced mode, the same content lives under the systemPrompt JSON key.

What it controls​

  • Persona — the agent's name, tone, and personality.
  • Scope — what the agent should and should not help with.
  • Style — reply length, formality, language preference.
  • Tool usage — when to call which tool.
  • Output format — bullets, JSON, code, plain text.

Auto-suggest​

After you type an Agent Name and click out of the name field, the dashboard may generate a draft system prompt suggestion. If it does, the suggestion appears as placeholder text in the System Prompt field. Click into the System Prompt field and press Tab to accept and insert the suggestion — you can then edit it from there.

If you've already typed something in the System Prompt field, the auto-suggest does not overwrite it.

Writing a good system prompt​

A good prompt is specific, short, and honest about limits. Template:

You are <NAME>, a <ROLE> for <AUDIENCE>.

You can:
- <capability 1>
- <capability 2>

You cannot:
- <restriction 1>
- <restriction 2>

Style:
- <tone, length, language>

When you need to <do X>, use the <TOOL NAME> tool.

Example​

You are Aria, a helpful assistant for Espressif customers.

You can:
- Answer questions about ESP devices using the documentation tool.
- Walk users through device setup step by step.
- Suggest related products from the catalog tool.

You cannot:
- Discuss internal pricing or unreleased products.
- Promise specific delivery dates.

Style:
- Friendly, concise. Reply in the user's language.
- Use bullet lists for step-by-step instructions.

If the user asks about a product, prefer the catalog tool over the docs tool.

Tips​

  • Be precise. "Reply in two sentences" beats "Reply briefly."
  • Tell the agent what NOT to do. Vague prompts produce vague behaviour.
  • Reference tools by name. The LLM chooses tools based on their name and description — nudging it in the prompt helps when two tools sound similar.
  • Don't dump every edge case in the prompt. Long prompts cost tokens on every turn. Tools and Knowledge Bases handle most edge cases better.
  • Iterate. Edit, test in a fresh conversation, edit again. Five minutes of prompt tuning often beats a model upgrade.

What NOT to put in a system prompt​

  • Secrets (API keys, passwords) — they are sent to the LLM provider on every turn.
  • Per-user personalization — end users have their own preference settings.
  • Long lists of facts — use a Knowledge Base instead.

On this page