Writing System Prompts
The system prompt is the most important part of any agent. It tells the AI who it is, what it should do, how it should communicate, and what it should avoid. A well-written system prompt produces consistent, useful agents. A vague one produces inconsistent, frustrating ones.
Character limit
System prompts are capped at 8,000 characters (roughly 1,500–2,000 words). This is usually far more than you need. Most high-quality system prompts are 200–500 characters.
The four elements of a good system prompt
1. Persona
Establish who the agent is:
"You are Alex, a technical support specialist for Acme Software. You are patient, precise, and friendly."
Giving the agent a name and a role helps it stay consistent and makes interactions feel intentional.
2. Tone and style
Specify how the agent should communicate:
"Reply in plain English. Avoid jargon. Keep responses under 150 words unless the user asks for detail. Use bullet points for steps."
3. Constraints
Tell the agent what it should not do:
"Do not discuss pricing — direct the user to the sales team. Do not speculate about competitors. If you don't know the answer, say so."
Never put real API keys, passwords, or sensitive data in a system prompt. System prompts are stored in plaintext and are visible to anyone with editor access to the agent.
4. Examples (optional but powerful)
A few examples of ideal responses help anchor the agent's style:
"When a user asks for a workaround, respond like this: 'One option is to [step]. Another is to [step]. Let me know which works for you.'"
Welcome message vs system prompt
These are different things:
| System Prompt | Welcome Message | |
|---|---|---|
| Visible to user? | No | Yes |
| Purpose | Instructions to the AI | First message shown to the user |
| Where configured | System Prompt tab | Overview tab |
The welcome message is what the agent says when a chat opens. The system prompt is what the agent knows and follows throughout the entire conversation.
Tips for iteration
- Start minimal — begin with 3–4 sentences and test. Add constraints only when you observe problems.
- Be specific about failures — if the agent gives answers that are too long, add "Keep responses under 100 words" rather than vague direction like "be concise".
- Use the test pane — run test cases in the builder's test pane before sharing the agent with teammates.
- Version history — every save creates a new version. If a change makes things worse, restore the previous version from the Versions menu.
If you want to give the agent company-specific knowledge (product docs, FAQs, policies), use the Knowledge tab to attach documents rather than pasting that content into the system prompt. RAG is more efficient and doesn't eat into your 8,000 character limit.
Was this helpful?