OneClickClaw.io
$

System Prompts & Personality

What is a system prompt?

A system prompt is a set of instructions that defines your agent's personality, behavior, and boundaries. It is the first thing the AI model reads before responding to any message, and it shapes every interaction your agent has. Think of it as the agent's job description: it tells the AI who it is, how it should behave, what topics it should focus on, and what it should avoid.

Without a system prompt, your agent behaves like a generic AI assistant. With a well-crafted system prompt, it becomes a specialized tool that reflects your brand, answers domain-specific questions, and maintains a consistent tone across all conversations.

On OpenClaw, these instructions live in your agent's workspace files. You edit them in the Workspace Files card on your server page. The main ones are:

Server page - Workspace Files (OpenClaw)

SOUL.md

Defines your agent's personality, values, and general vibe.

IDENTITY.md

Who your agent is when speaking to people: name, role, what it represents.

USER.md

What your agent should remember about you: how to address you, your preferences, your context.

AGENTS.md

Lists the helper agents your main agent can hand work to, and when to use each one.

On Hermes Agent, you edit the same kind of instructions inside your Hermes dashboard, which you open with Open Agent.

Writing effective prompts

A good system prompt covers five areas. You do not need to include all five in every prompt, but addressing them produces the most consistent results:

  • Role definition. Tell the agent exactly who it is. A vague line like "be helpful" gives it little to work with. Write something like: "You are the customer support assistant for Acme Software. You help customers with billing, account issues, and product questions."
  • Topic boundaries. Define what the agent should and should not discuss. For example: "Do not provide medical advice. If someone asks a medical question, recommend they consult a professional."
  • Tone and style. Specify whether the agent should be formal, casual, friendly, professional, or humorous. For example: "Use a warm, professional tone. Avoid jargon. Keep sentences short."
  • Audience context. Describe who talks to the agent so it adjusts its language level. For example: "Your users are small business owners who are not technical. Explain things simply."
  • Response format. If you want short answers, say so. If you want bullet points, numbered steps, or a specific structure, include that instruction.

Here are four practical examples for common use cases:

Example: Customer support bot

Note

"You are the customer support assistant for Acme Software. Your job is to help customers with billing questions, account issues, and product features. Be polite and professional. Keep responses concise (under 200 words). If you do not know the answer, say so honestly and suggest the customer contact support@acme.com for further help. Never make up information about pricing or policies. Always verify account details before making changes."

Example: Sales assistant

Note

"You are a sales assistant for GreenTech Solar Panels. Your goal is to answer questions about our residential solar products, help potential customers understand pricing and savings, and guide them toward booking a free consultation. Be enthusiastic but not pushy. Use concrete numbers when discussing energy savings. If someone asks about commercial installations, explain that we focus on residential and offer to connect them with our commercial team. Never discuss competitor products negatively."

Example: Community moderator

Note

"You are the community moderator for the Pixel Art Discord server. You help members with pixel art techniques, tool recommendations, and creative feedback. Be enthusiastic and encouraging. Use casual language. When members share their work, offer constructive feedback with specific suggestions for improvement. If you see heated arguments, gently redirect the conversation. Do not engage in off-topic debates, controversial discussions, or political topics. Remind members of server rules when appropriate."

Example: Knowledge base assistant

Note

"You are the knowledge base assistant for CloudDrive file storage. You answer questions about features, pricing plans, file sharing, sync settings, and troubleshooting. Use short, direct answers organized with bullet points. If a question requires a multi-step solution, use numbered steps. Reference our documentation when possible. If you cannot find the answer, suggest the user search our help center at help.clouddrive.com or open a support ticket. Never speculate about upcoming features or share internal information."

Best practices and tips

  • Iterate and refine. Your first prompt will probably not be perfect. Test it with real conversations, notice where the agent falls short, and adjust the prompt accordingly. Small changes often have a big impact on response quality.
  • Include examples in the prompt. If you want a specific response format, show the agent an example: "When asked about shipping, respond like this: Your order will arrive in 3 to 5 business days. Express shipping is available for an additional EUR 9.99."
  • Keep it focused. A prompt that tries to cover everything often produces unfocused responses. Stick to the agent's core purpose. Prompts under 2000 characters tend to produce the most consistent behavior.
  • Plan for edge cases. What happens if someone asks something off-topic? What if someone tries to manipulate the agent? Add instructions for how the agent should handle these situations gracefully.
  • Use "do not" instructions sparingly. Telling the agent what to do works best. For example, write "Respond with warmth and patience, even to frustrated customers."

Tip

Keep your system prompt under 2000 characters. Shorter, focused prompts produce more consistent behavior than long, detailed ones. The AI model pays the most attention to the first few sentences, so put your most important instructions at the top.

How to update your system prompt

1

Open the server page

On the dashboard overview, click Control Panel on your OpenClaw server card.

2

Open the Workspace Files card

Scroll to Workspace Files and select a file, for example SOUL.md for personality. Its content opens for editing. About these files explains what each one does. The Files link at the top of the server page opens the same files on a page of their own.

3

Edit and save

Change the text and save the file. Your agent reads its workspace files when it works, so send it a new message to see the change.

On Hermes Agent, click Open Agent and edit the instructions inside your Hermes dashboard.

Testing your prompt changes

After updating your system prompt, test it thoroughly before considering it done:

  • Send a few typical questions your users would ask and verify the tone and accuracy.
  • Try an off-topic question to confirm the agent handles it according to your boundaries.
  • Ask something the agent should not answer (if you set restrictions) and verify it declines appropriately.
  • Test a long, multi-part question to see if the agent stays coherent and follows your formatting preferences.
  • Have someone else interact with the agent without telling them the prompt. Fresh eyes catch issues you might miss.

Tip

Keep a changelog of your system prompt edits. Note what you changed and why. This makes it easy to roll back if a change does not work out, and helps you understand which instructions have the biggest impact on your agent's behavior.

Note

For deeper guidance on choosing the right model to pair with your prompt, see Choosing the Right Model. Different models respond differently to the same prompt, so model choice and prompt writing go hand in hand.