A job description for your AI agent.
Before you hand a job to AI, write down exactly what it is, who owns it, what it may touch, what it must never do, and how you will know it is working. One page, nine fields. It’s free and yours to keep, and it works with whichever AI you use. You don’t need to give us an email address.
Nine questions the tutorial videos skip.
Writing the plan is quick. Answering these honestly is the part that decides whether handing the job over is a good idea.
| Field | The question | The rule |
|---|---|---|
| Job | What exactly is being handed over? | One line, in your own words. One job per card. |
| Owner | Who is accountable when it goes wrong? | A named person and their role ("Sam, office manager"), never "the AI". |
| Payoff | Why is it worth doing? | Time, money, a problem solved or a goal served, in your words. Only use numbers you actually know. |
| Inputs | What may it read? | Name the tools and the data. Everything else is off-limits by default. |
| Outputs | What does it produce? | Mark each one "drafts for approval" or "acts on its own". |
| Boundary | What must it never do? | Anything involving money, passwords or messages to customers needs approval by default. Add your own "nevers". |
| Check | How will you know it worked? | Something you can see, like "every quote gets a follow-up within 2 days". |
| Review date | When do you look again? | A date two to six weeks away. |
| Verdict | Is it worth handing over? | Worth it, Not yet, or Not worth it, with the reason. "Not worth it" is a valid, finished answer. |
Quote follow-ups for a landscaping business.
A made-up example, to show what a finished card looks like.
| Job | Follow up every quote we send that hasn’t had a reply. |
| Owner | Jess, office manager |
| Payoff | Quotes go cold because nobody chases them. |
| Inputs | The scheduling tool’s quote list; the shared email inbox (read only). |
| Outputs | Follow-up email: drafts for approval. Weekly list of cold quotes: acts on its own (internal only). |
| Boundary | Never sends to a customer without approval. Never changes a price. Never touches invoices or payments. |
| Check | Every quote over 3 days old has a drafted follow-up waiting for Jess. |
| Review date | Four weeks from the start. |
| Verdict | Worth it. Low risk while it only drafts, and the result is easy to see. |
Type it here, keep it as a PDF.
Fill in the fields, then print it or save it as a PDF from your browser. Nothing is sent to us.
Let your AI interview you.
Paste this into Claude, ChatGPT, Copilot or Gemini. It asks you one question at a time and fills in the card, including an honest “not worth it” if that is the answer.
You are helping me write a Role Card: a one-page job description for an AI agent. Interview me about ONE job in my business, one question at a time. Fill in these fields: Job, Owner, Payoff, Inputs, Outputs, Boundary, Check, Review date, Verdict. Rules: Use my own words for Job and Payoff. Never invent numbers I didn't give you. The Owner must be a named person, never "the AI". Mark every output as "drafts for approval" or "acts on its own". Anything involving money, passwords or messages to customers defaults to "needs approval". Only list tools I said I use. For Verdict, choose Worth it / Not yet / Not worth it and give an honest reason. Say "not worth it" if that's the truth. When finished, show the completed card as a table.
Writing it down doesn’t prove it’s true.
This card is a plan. It doesn’t check that your AI actually behaves this way. That’s the Agent Readiness Review. We check every field against your real system: are the inputs really the only access, are the outputs really drafts, can the boundary be got around. $1,200 incl. GST, 10 business days, one agent or one workflow.
Writing down what is allowed and then checking it with fixed rules is how we work generally. Watch shows that method running in public, on open-source packages rather than anyone’s agent.