System Prompts

Prompt Structure

A good system prompt has four parts:

1. Role Definition

Tell the agent what it is and what it does.

You are a tech news analyst that creates daily digest summaries for startup founders.

2. Output Format

Define how the agent should structure its responses.

Respond in markdown with:
- A one-line summary
- Key insights as bullet points
- A "Why it matters" section for each story

3. Boundaries

Set what the agent should and should not do.

Only discuss technology, startups, and venture capital topics.
Do not provide investment advice.
If asked about unrelated topics, politely redirect.

4. Tool Instructions

If the agent has MCP tools, explain when and how to use them.

Use the search_memory tool to recall previous conversations.
Use the web_fetch tool to retrieve current information when asked about recent events.
Always cite your sources when using fetched data.

Best Practices

• Be specific — “You are a Python code reviewer” beats “You are helpful”
• Include examples — Show the agent what good output looks like
• Set tone — Use the responseTone field (professional, casual, friendly, formal, technical, creative)
• Limit scope — Agents that try to do everything do nothing well
• Reference variables — Use {{input.message}} to reference flow variables in prompts

Temperature Guidelines

Task Type	Temperature	Example
Factual / analytical	0.1–0.3	Code review, data extraction
Balanced	0.4–0.6	Customer support, Q&A
Creative	0.7–0.9	Writing, brainstorming
Experimental	1.0+	Idea generation