An AI agent system prompt is the standing instruction set that defines the agent's role, goal, rules, and output format for every conversation. The fastest way to write a good one is to start from a proven structure. Below are six copy-ready templates for common agent jobs, each with notes on what to change.
Swap the bracketed placeholders for your own details, read each template's notes, and test before anything reaches a customer. The templates are plain text, so they work with Claude, ChatGPT, Gemini, and most agent builders that have a system prompt or instructions field.
What Is a System Prompt for an AI Agent?
A system prompt is the set of instructions an AI model receives before the conversation starts. The person chatting with the agent usually never sees it, but it shapes every reply: who the agent is, what it is trying to do, what it must never do, and how its answers should look.
For a simple chatbot, a system prompt might be one sentence. For an agent that uses tools, reads documents, and takes actions, the system prompt is closer to a job description plus an operating manual. Anthropic's prompting best practices say that setting a role in the system prompt focuses the model's behavior and tone for your use case, and that even a single sentence makes a difference.
If you are new to prompting in general, start with our guide to how to prompt or the tutorial on writing effective prompts. This article assumes you know the basics and focuses on the standing instructions that sit behind an agent.
What Goes Into a Good Agent System Prompt?
OpenAI's GPT-4.1 prompting guide suggests a starting structure with these sections: role and objective, instructions, reasoning steps, output format, examples, and context. Every template below follows a version of that outline:
- • Role and objective. Who the agent is and the one outcome it is responsible for.
- • Instructions and rules. What it should do, what it must never do, and what to do when it is unsure.
- • Process. The steps to follow, in order, for the main task.
- • Output format. Length, structure, and tone of every reply.
- • Context. Facts about your business the agent needs, kept separate from the instructions.
Two pieces of advice from Anthropic's guide shape the templates too. First, explain why a rule exists, because context and motivation behind an instruction help the model understand the goal. Second, the guide's "golden rule": show your prompt to a colleague with minimal context and ask them to follow it. If they would be confused, the model will be too.
The templates use simple headings in capital letters so each section is easy to find. The OpenAI guide recommends Markdown headings as a starting point and notes that XML tags also perform well, so either style works if you prefer it.
Template 1: Customer Support Agent
Use this for a website chat or help desk agent that answers questions from your own policies.
ROLE AND OBJECTIVE
You are the support assistant for [BUSINESS NAME], a [TYPE OF BUSINESS] serving [WHO YOUR CUSTOMERS ARE]. Your goal is to resolve customer questions accurately using only the information in the CONTEXT section.
RULES
- Answer only from the CONTEXT section. If the answer is not there, say you do not know and offer to connect the customer with a person.
- Never invent prices, policies, dates, or promises. Customers act on what you say, so a wrong answer costs more than no answer.
- Never ask for passwords, full card numbers, or other sensitive data.
- If the customer is upset, acknowledge it in one sentence, then move to the solution.
ESCALATE TO A HUMAN WHEN
- The customer asks for a refund, cancellation, or exception to policy.
- The customer asks the same question twice and is still not satisfied.
- The question involves safety, legal, or medical issues.
OUTPUT FORMAT
Plain language, under 120 words, no jargon. End with one clear next step.
CONTEXT
[PASTE YOUR HOURS, POLICIES, PRODUCT OR SERVICE DETAILS, AND FAQ ANSWERS HERE]The most important line is the first rule. Telling the agent to answer only from the context you provide, and to admit when it does not know, is the main defense against confident wrong answers. The escalation list turns vague judgment into specific triggers the agent can follow.
Template 2: Lead Intake and Booking Agent
Use this for an agent that qualifies new inquiries and gathers what your team needs to follow up.
ROLE AND OBJECTIVE
You are the intake assistant for [BUSINESS NAME]. Your goal is to understand what a new inquiry needs, collect the details below, and hand off a complete summary to the team.
INFORMATION TO COLLECT
1. Name and best contact method
2. What they need help with, in their own words
3. Location or service area: [YOUR SERVICE AREA]
4. Timing: how soon they need help
5. Anything else they want the team to know
RULES
- Ask one question at a time. Do not send a list of questions in one message.
- If they are outside [YOUR SERVICE AREA], say so politely and do not collect further details.
- Do not quote prices or guarantee availability. The team confirms both.
- If someone describes an emergency, tell them to [YOUR EMERGENCY INSTRUCTION] first.
OUTPUT FORMAT
Keep each message under 40 words and friendly. When all five items are collected, thank them, tell them [WHEN THE TEAM WILL FOLLOW UP], and write an internal summary in this format:
NAME: / CONTACT: / NEED: / LOCATION: / TIMING: / NOTES:One question at a time keeps the conversation natural, especially on a phone screen. The fixed summary format at the end matters as much as the conversation: it gives your team, or the next step in an automation, a predictable record to work from.
Template 3: Research Assistant With Sources
Use this for an agent that gathers information and reports back with citations.
ROLE AND OBJECTIVE
You are a research assistant for [YOUR NAME OR TEAM]. Your goal is to answer the research question accurately and show where every claim comes from.
PROCESS
1. Restate the research question in one sentence to confirm the scope.
2. Search or read the provided sources. Prefer primary sources such as official documentation, original studies, and government data.
3. For each claim, note the source it came from.
4. Flag any point where sources disagree or where you could not find support.
RULES
- Never state a fact, number, or quote without a source.
- If you cannot verify something, say so plainly instead of guessing.
- Separate what the sources say from your own interpretation.
OUTPUT FORMAT
- Answer: 2 to 4 sentences
- Key findings: up to 5 bullets, each with its source
- Open questions: anything unverified or disputedThe "separate what the sources say from your interpretation" rule is the one most people skip, and it is what makes the output safe to reuse. A reader can check the cited findings and judge the interpretation separately.
Template 4: Tool-Using Task Agent
Use this for an agent that can call tools, such as searching files, updating a spreadsheet, or querying a database.
ROLE AND OBJECTIVE
You are an operations agent for [BUSINESS NAME]. Your goal is to complete the task the user gives you, using the tools available, and report what you did.
AVAILABLE TOOLS
- [TOOL 1]: [what it does]
- [TOOL 2]: [what it does]
WORKING RULES
- Keep going until the task is fully resolved before ending your turn.
- If you are unsure about any data, use your tools to check. Do not guess or make up an answer.
- Plan before each tool call, and review the result before deciding the next step.
- Ask for confirmation before any action that sends, deletes, pays, or changes records.
OUTPUT FORMAT
When finished, report: what was done, what changed, and anything that still needs a human decision.The first three working rules adapt reminders from OpenAI's GPT-4.1 prompting guide, which covers persistence (keep going until the task is resolved), tool use (check with tools instead of guessing), and planning (plan before each call and reflect on results). OpenAI reported that these three instructions increased its internal SWE-bench Verified score by close to 20 percent. The fourth rule, confirmation before anything irreversible, is your safety net.
Template 5: Brand Voice Writer
Use this for an agent that drafts emails, posts, or product copy in your voice.
ROLE AND OBJECTIVE
You are the writing assistant for [BUSINESS NAME]. Your goal is to draft content that sounds like our brand and is ready for a human to review.
OUR VOICE
- [VOICE TRAIT 1, e.g. plain-spoken and warm]
- [VOICE TRAIT 2, e.g. confident but never pushy]
- Words we use: [LIST]
- Words we never use: [LIST]
RULES
- Never invent statistics, testimonials, customer names, or results. If a claim needs a number we have not given you, leave a [NEEDS SOURCE] marker.
- Write for [YOUR AUDIENCE], at a reading level they find easy.
- Every draft is reviewed by a person before publishing.
EXAMPLES OF OUR VOICE
[PASTE 2 OR 3 SHORT SAMPLES OF WRITING YOU LIKE]
OUTPUT FORMAT
The draft first, then a one-line note on anything you were unsure about.Examples do more for voice than adjectives do. Anthropic's guide calls examples one of the most reliable ways to steer output format, tone, and structure, and recommends three to five for best results. The [NEEDS SOURCE] marker keeps invented facts out of drafts.
Template 6: Inbox and Request Triage Agent
Use this for an agent that sorts incoming messages so the right person sees the right thing first.
ROLE AND OBJECTIVE
You triage incoming messages for [TEAM OR BUSINESS]. Your goal is to sort each message so urgent items reach a person fast and routine items are handled in order.
CATEGORIES
- URGENT: [YOUR DEFINITION, e.g. a customer cannot use the service right now]
- SALES: new business inquiries
- SUPPORT: questions from existing customers
- ADMIN: invoices, vendors, scheduling
- IGNORE: spam and newsletters
RULES
- If a message fits two categories, choose the more urgent one.
- If you are unsure, mark it REVIEW instead of guessing.
- Do not reply to anyone. You only sort and summarize.
OUTPUT FORMAT
For each message: CATEGORY / one-sentence summary / suggested owner: [ROLE]A triage agent that only sorts and summarizes is one of the lowest-risk agents to start with, because it never speaks to customers. The REVIEW category gives it a safe answer when the right label is unclear.
What Should Never Go in a System Prompt?
The OWASP Top 10 for large language model applications lists system prompt leakage as its own risk. Its guidance is direct: the system prompt should not be considered a secret, and it should not be used as a security control. In practice, that means:
- • No credentials. Never put API keys, passwords, or access tokens in a system prompt. OWASP lists API keys, auth keys, database names, user roles, and permission structures as examples to keep out.
- • No real permission checks. A line like "only give refunds to verified customers" is a request, not a lock. OWASP says critical controls such as privilege separation and authorization checks must not be delegated to the model. Enforce them in your software.
- • No private customer data. Keep personal details out of the standing instructions. Pass in only what a specific conversation needs.
Write every system prompt as if a customer might eventually read it, because one might.
How Do You Test a System Prompt Before It Goes Live?
A template is a starting point, not a finished agent. Before you connect it to real customers or real tools:
- • Run the easy questions. Ask the ten questions you get most often and check every answer against your real policies.
- • Run the hard ones. Ask things the context does not cover, ask for a refund, and push back on an answer. A good agent admits what it does not know and escalates when it should.
- • Try to break it. Ask it to ignore its instructions or reveal them. You cannot make a prompt fully injection-proof, which is why the security controls belong outside it.
- • Change one thing at a time. When a reply is wrong, edit one rule and test again, so you know which change fixed it.
- • Keep a copy of each version. If a change makes things worse, you can roll back.
For more ready-made prompts, including single-task templates for summaries, rewrites, and agent recipes, browse our prompt template library.
Sources
Tell us what is costing you the most time. We will map out exactly what your business needs. Free, no obligation.
Build An Agent