AI Agent System Prompt Template: The 6 Parts Behind Every BotPlaybooks Persona

Updated: November 12, 2026 · By Brian Arrington (Founder, AI Consultant and Business Coach)*

The short answer

A good AI agent system prompt has six parts: the role, the inputs it gets, the workflow it follows, the output format, the guardrails, and an example first task. Copy the template below, fill in each part for one job, and test it with the first task before you add any tools.

Most agents that disappoint people were never told what their job is. They got a name, a friendly tone and a hope. The fix is not a longer prompt. It is a clearer one, with the same six parts every time.

That is the structure behind every persona in the BotPlaybooks library. Below is the empty skeleton, free to copy, followed by each part explained with real public fields from two of our personas: the agent called "Chief" (Chief of Staff) and the agent called "Docket" (Inbox Triage Assistant).

The template (copy and fill in)

Paste this into a plain text file and fill in each section for one job. Keep the headings. They make the prompt easy to read for you and for the agent.

# [Agent name]: [Job title]

## Role
Who this agent is and the one job it owns, in one or two sentences.

## Inputs
What you will give it: data, files, accounts and access level (start read-only).

## Workflow
The steps it follows, in order. Say what to do when information is missing.

## Output format
What it hands back and in what shape: a brief, a table, a checklist, a draft.

## Guardrails
What it must ask before doing, and what it never does.

## Example first task
One ready message you can send in a new chat to test it.

That is the whole template. The rest of this guide shows what good looks like under each heading.

Part 1, Role: one agent, one job

Start with a name and a single job. A role line should be short enough to read aloud and specific enough that the agent can tell when a request is outside it.

Chief example: "Runs the operating rhythm for a leader: priorities, meeting prep, follow-through and decisions."

Notice what that line does not say. It does not say "helps with anything" or "is a smart assistant." It names an owner (a leader), a rhythm and four areas. When a request lands outside those, the agent has a reason to ask before it acts.

Tip: one job per agent. If the role needs the word "and" more than once to join two different jobs, you probably need two personas. An inbox agent and a sales agent can share an account. They should not share a prompt.

What should an AI agent system prompt include? It should include six parts: a role with one job, the inputs the agent gets, a step-by-step workflow, the output format, guardrails for what needs approval, and one example first task. Tools, connectors and routines are set up around the prompt in your agent tool.

Part 2, Inputs: what you give it

List the data and access the agent needs to do its job. Be concrete. "My stuff" is not an input. "My calendar for the next 7 days, read-only" is.

Chief example: "your calendar, inbox (read), goals or OKRs, meeting notes, and who owns what."

Docket example: "inbox access, the team chat channels and direct messages I approve, my priorities, my VIP list, and whether I want full triage or the top-3 brief."

Docket's list is a good model because it includes a choice (full triage or a brief) as an input. Telling the agent which mode you want is part of what you give it.

Start with the least access that works. Grok Bot's own use cases page says it well: "Every example stops for your review. Keep sending, publishing, purchasing, and production changes behind approval." [3] Read access first. Write access later, once you trust what comes back.

Part 3, Workflow: the steps in order

The workflow is where most prompts go vague. Write the steps as a numbered list, in the order you would do them yourself. Here is a generic example for a weekly report agent (this is an illustration, not the workflow inside any BotPlaybooks persona):

1. Read the inputs listed above for the date range I give you.
2. Pull out what changed since last week.
3. Group the changes under the headings in the output format.
4. Flag anything that needs my decision at the top.
5. If a number, date or name is missing, mark it "[missing]" and list it at the end. Never guess.
6. Hand back the draft and stop.

Step 5 matters more than the rest. An agent that is told what to do when information is missing will ask or mark it. An agent that is not told will fill the gap, and that is where made-up facts come from.

Part 4, Output format: what it hands back

Name the deliverable and its shape. "A summary" is loose. "A one-page brief with three headings: decisions needed, today's meetings, open follow-ups" is something you can check.

Chief example: "A daily brief, meeting agendas written as decisions to make, a decisions log, an action tracker with owners and due dates, and a weekly summary against your goals."

Docket example: "In full mode, a summary, a categorized list across email and chat, draft replies and extracted tasks. In brief mode, only the top 3 urgent items with source links and one suggested next step each."

Both examples name things you would recognize on your desk. That is the test. If you cannot picture the output, the agent cannot either.

Part 5, Guardrails: the approval line

Guardrails say two things: what the agent must ask before doing, and what it never does. This is the part people skip, and it is the part that keeps an agent useful instead of risky.

Chief and Docket share the same guardrail: "Drafts only: it never sends, posts, pays or contacts anyone without your approval, and never invents facts."

That single line is a complete ai agent guardrails example. It draws the approval line at anything that leaves your hands (sending, posting, paying, contacting) and it closes the door on made-up facts. Add job-specific lines under it, such as "Ask before archiving anything from my VIP list."

What are guardrails in an AI agent prompt? Guardrails are the rules that set the approval line. They list what the agent must ask before doing, such as sending an email or making a payment, and what it never does, such as inventing facts. Clear guardrails let you give an agent real work while you keep the final say.

Part 6, Example first task: test it in one message

The last part is a single, ready message you send in a new chat. It proves the prompt works before you spend time on tools or routines.

Chief example: "Act as my chief of staff this week. Start with tomorrow's brief, then set up a decision log and action tracker from my last three meeting notes."

Docket example: "Triage my inbox and my Slack DMs from the last 24 hours, then give me just the top 3 urgent items."

A good first task is small, real and easy to judge. You know what tomorrow looks like, so you can tell in a minute whether the brief is right. If it is not, fix the prompt, not the task.

Around the prompt: tools, connectors and routines

A persona file has two more sections that sit around the prompt: tools and connectors, and routines. These are not just words. You turn them on in your agent tool.

Chief's tools: "Calendar, email (read, drafts only), notes, task or project app, shared docs for the decision log." Docket's tools: email (read and draft), team chat such as Slack or Microsoft Teams (read, approved channels and DMs) and a task app.

Connect only what the inputs list needs. Add routines, such as a brief every weekday morning, only after you have reviewed a few outputs by hand and trust them.

System prompt vs user prompt

The system prompt is the standing job: the role, the workflow, the format and the rules that stay true in every chat. The user prompt is today's request, such as "Start with tomorrow's brief." OpenAI calls the standing part "developer messages" and says they "are instructions provided by the application developer, prioritized ahead of user messages," while user messages "are instructions provided by an end user." [10] Anthropic uses the term system prompt and says "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference." [11]

Put anything you would repeat every day in the system prompt. Put anything that changes day to day in the user prompt. If you keep typing the same instruction, it belongs in the system prompt.

Where each part goes in your tool

The six parts are the same everywhere. Where you paste them changes by tool. See the setup guide for full steps.

For OpenClaw, our SOUL.md templates for business agents show the split in detail.

How long should all this be? No major vendor sets one length for agent prompts. Microsoft's Azure OpenAI guide lists "overly long system messages" as a common pitfall, because "longer messages can consume context window and reduce room for user content." [12] Cover the six parts and stop there.

Skip the blank page

The template is free, and it works. If you would rather start from a prompt already written for your job, the BotPlaybooks library has 175+ personas built on these six parts, including the Chief of Staff persona and the Inbox Triage Assistant persona. Each one is one plain Markdown file.

Skip the blank page: get a persona already written for your job. Browse the packs by department, or get a free persona to try one first.

FAQ

What is an AI agent system prompt template?

It is a reusable outline for the standing instructions you give an AI agent. You fill in the same six headings for each job: role, inputs, workflow, output format, guardrails and an example first task. The structure stays the same, so each new agent is faster to write and easier to check.

What should an AI agent system prompt include?

It should include six parts: a role with one job, the inputs the agent gets, a step-by-step workflow, the output format, guardrails for what needs approval, and one example first task. Tools, connectors and routines are set up around the prompt in your agent tool, not only written into it.

How long should a system prompt be?

No major vendor sets one length for agent prompts. Microsoft warns that overly long system messages use up context and leave less room for your request. Cover the six parts clearly and stop there. If you keep repeating an instruction in chat, add it. If a line never changes the output, cut it.

What is the difference between a system prompt and a user prompt?

The system prompt, which OpenAI calls developer messages, is the standing job: the role, workflow and guardrails that stay true in every chat. The user prompt is the request you send today, such as asking for tomorrow's brief. Anything you would repeat every day belongs in the system prompt, not in each message.

Can I use the same system prompt in Grok Bot, Hermes and OpenClaw?

Yes. The six parts stay the same, and only the placement changes. In Grok Bot, use the plus button in your main Grok Bot to give it the Markdown file and tell it to build the persona as a new agent. In Hermes, put the persona (including standing approval rules, and any workflow that should follow the agent everywhere) in SOUL.md, and put only project-specific instructions in AGENTS.md. In OpenClaw, official docs keep voice, stance and style in SOUL.md and operating rules in AGENTS.md; you can paste the whole Persona / System Prompt into SOUL.md, or use that official split. Test it with the same first task in each tool.

Where can I get ready-made system prompts for business agents?

The BotPlaybooks library has 175+ personas for business and personal agents, each written with the same six parts and saved as one plain Markdown file. Each persona page shows the role, inputs, outputs, guardrails and an example first task, so you can see what you are getting before you choose one.

Related guides

• SOUL.md templates for business agents

AI can make mistakes. Always review outputs before acting. Personas are templates, not legal, financial, tax, medical or HR advice. BotPlaybooks is independent and not affiliated with Meta, OpenAI, xAI, Nous Research, OpenClaw or any agent platform.

Sources

Sources 1 to 8 as cited in blog 08 (read October 21, 2026). Sources 9 to 12 and persona pages read October 22, 2026. Recheck all on publish day.

1. Grok Bot, create and manage Bots: https://docs.x.ai/grok-bot/bots
2. Grok Bot, overview: https://docs.x.ai/grok-bot/overview
3. Grok Bot, use cases: https://docs.x.ai/grok-bot/use-cases
4. BotPlaybooks, Chief of Staff persona: /personas/chief-of-staff
5. Hermes Agent, personality and SOUL.md: https://hermes-agent.nousresearch.com/docs/user-guide/features/personality
6. OpenClaw, SOUL.md personality guide: https://github.com/openclaw/openclaw/blob/main/docs/concepts/soul.md
7. Meta, Muse Identity, Soul and Memory files: https://www.meta.com/help/artificial-intelligence/2225571704857152/
8. OpenAI, Dots: https://learn.chatgpt.com/docs/dots
9. BotPlaybooks, Inbox Triage Assistant persona: /personas/inbox-triage-assistant
10. OpenAI, text generation guide (developer and user messages): https://platform.openai.com/docs/guides/text
11. Anthropic, giving Claude a role with a system prompt: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/system-prompts
12. Microsoft Learn, Azure OpenAI system message guidance: https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/advanced-prompt-engineering

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What is an AI agent system prompt template?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It is a reusable outline for the standing instructions you give an AI agent. You fill in the same six headings for each job: role, inputs, workflow, output format, guardrails and an example first task. The structure stays the same, so each new agent is faster to write and easier to check."
      }
    },
    {
      "@type": "Question",
      "name": "What should an AI agent system prompt include?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It should include six parts: a role with one job, the inputs the agent gets, a step-by-step workflow, the output format, guardrails for what needs approval, and one example first task. Tools, connectors and routines are set up around the prompt in your agent tool, not only written into it."
      }
    },
    {
      "@type": "Question",
      "name": "How long should a system prompt be?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "No major vendor sets one length for agent prompts. Microsoft warns that overly long system messages use up context and leave less room for your request. Cover the six parts clearly and stop there. If you keep repeating an instruction in chat, add it. If a line never changes the output, cut it."
      }
    },
    {
      "@type": "Question",
      "name": "What is the difference between a system prompt and a user prompt?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The system prompt, which OpenAI calls developer messages, is the standing job: the role, workflow and guardrails that stay true in every chat. The user prompt is the request you send today, such as asking for tomorrow's brief. Anything you would repeat every day belongs in the system prompt, not in each message."
      }
    },
    {
      "@type": "Question",
      "name": "Can I use the same system prompt in Grok Bot, Hermes and OpenClaw?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Yes. The six parts stay the same, and only the placement changes. In Grok Bot, use the plus button in your main Grok Bot to give it the Markdown file and tell it to build the persona as a new agent. In Hermes, put the persona (including standing approval rules, and any workflow that should follow the agent everywhere) in SOUL.md, and put only project-specific instructions in AGENTS.md. In OpenClaw, official docs keep voice, stance and style in SOUL.md and operating rules in AGENTS.md; you can paste the whole Persona / System Prompt into SOUL.md, or use that official split. Test it with the same first task in each tool."
      }
    },
    {
      "@type": "Question",
      "name": "Where can I get ready-made system prompts for business agents?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The BotPlaybooks library has 175+ personas for business and personal agents, each written with the same six parts and saved as one plain Markdown file. Each persona page shows the role, inputs, outputs, guardrails and an example first task, so you can see what you are getting before you choose one."
      }
    }
  ]
}