How Do I Write an AI Agent Job Description So It Knows What to Own, What to Ask, and What to Hand Back to Me?

An AI agent job description has three parts: what the agent owns end to end, the exact conditions where it stops and asks you, and what it hands back when the work is done. Write those three things in plain English and your agent goes from answering questions to finishing work. Skip them and you get another chat that starts from zero every morning.

Why does a prompt stop working where a job description keeps working?

I hire a lot. Mostly agents these days, though I have hired plenty of people too. I have watched the same mistake show up in almost every business I look at, including my own early attempts at this.

People write prompts. Prompts are tasks. "Summarize this email." "Write me a follow up." The task finishes, the agent goes back to sleep, and tomorrow you type the whole context again.

A job description is a standing role. It says: this is the territory you own, this is your authority, these are the edges, and this is what you bring back to me. That is the jump from busy to useful. Busy is an agent that answers 40 questions a day. Useful is an agent that closes the loop on inbound leads while I am standing in a wedding venue on a Saturday.

The bottleneck is almost never the model. It is the description. I have trained more than 90,000 people on software and systems since 1998, and the pattern has not changed: the tool was never the problem. The instructions were.

What should your agent own without asking me?

Ownership means a whole loop, not a step.

Take inbound leads. An agent that owns the loop watches the form, checks whether the person is already in the CRM, pulls the last three emails, writes the first reply draft, tags the record, and schedules the follow up. End to end. If I have to touch it in the middle, it does not own that loop. I either split the loop into two roles or hand it to a person.

Then write the boundary in the same breath. "You own inbound leads from form submission through first reply draft. You do not own pricing. You do not own refunds. You do not touch accounts I have had for more than a year."

For the wiring I use n8n or Make.com on the trigger, Airtable or Notion as the record, and Claude or ChatGPT for the judgment inside the loop. All of it no code. You do not need an engineer to get this running. You need one clearly bounded loop that happens at least five times a week, because that is where agentic workflows start paying for themselves instead of just looking clever.

What should your agent ask you about?

"Ask if unsure" is not an instruction. It is a shrug.

Name the triggers instead. Mine look like this: any spend over $500, anything that commits the company to a date, anything with legal exposure, anything going to a client I have known more than a year, and anything where the answer is not in the files you gave it.

Then define the shape of the ask. Four lines: what I tried, what I need from you, my recommendation, my confidence. That format takes me 20 seconds to answer instead of 10 minutes to untangle. An agent that dumps its confusion on you is just as expensive as one that guesses wrong.

This is also the real way to teach AI your business. Every question you answer is training data. I keep a running file of my answers, and when the same question gets answered the same way three times, I move it out of the ask column and into the owns column. The agent grows without me rewriting the whole thing. Founders and operators who skip this step end up re-explaining their business forever, to every agent, in every chat.

What should your agent hand back to me?

Hand back a deliverable, not a transcript. Nobody wants 900 words of reasoning.

Give it a format. Three lines: what happened, what I did, what I need from you. Then the artifact itself, the draft, the updated record, the list, the file. And define done in plain words. "Done means the lead is tagged, the reply draft is sitting in drafts, and the follow up is on the calendar." Without that definition, the agent keeps going and you get a status report every hour until you turn it off.

How do you write and test one this week?

Six steps, in order.

1. Pick one loop that happens at least five times a week.
2. Write the role in a plain document with three headers: OWN, ASK, HAND BACK.
3. Paste it into a Claude Project or a custom GPT along with your context files: what you sell, who buys it, how you talk, and three examples of work you are proud of.
4. Run ten real historical examples through it.
5. Log every time it did something you would not have done. That log is your edit list.
6. Rewrite the description, not the prompt.

When the description survives a week, then wire it up. n8n, Make.com, or Zapier for the plumbing. Airtable or Notion for the record. Claude or ChatGPT for the thinking. I run 45 agents now and every one of them started as a paragraph in a document before it ever touched an automation.

FAQ

How long should an AI agent job description be?

Mine run 300 to 600 words. Long enough to name the edges, short enough to read in two minutes. If it goes past a page, I usually find two jobs hiding in it, and I split them into two agents.

Can I write one without any coding?

Yes. Write the description first and test it by hand in a Claude Project or a custom GPT. Only automate the parts that survive real examples. Most of the value shows up before you connect anything.

How often should I update it?

Every time you answer the same question twice. I do a real pass once a month on any agent that touches money or clients, and everything else gets edited when the log tells me to.

I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book. And if you want to tell me which loop you are handing over first, I would love to hear it. That is usually where the interesting part starts.