What Should I Write in My AI Agent's Instructions?
In the session I used three words and the AI agent wrote its own. The file on the session page shows what a full set looks like, one section at a time.
Write the instructions in sections: who the AI agent is, what it does, what it reads, what it never does and how it takes a correction. In the session I asked for them with three words, "update agent instructions", and the AI agent wrote the first set. The file on the session page shows a full version.
You can type your own instructions or let the AI agent write them. I did the second one on October 7, 2026, and then I opened the box and named what it had written. The file on the session page is a full set for the Chief of Staff AI agent, so you can see what each section is for and write your own.
Do I have to write the instructions myself?
No. When I reached the instructions box I told the group you can type your own, and I also told them what happens if you let the builder do it.
From the recording, 23:51. You'll notice here, there's a section that says add instructions for agents. So if you're familiar with setting up agents, you can type in your own instructions.
From the recording, 24:07. So the information that we provided in the first screen and then the questions that we've answered, it will create instructions for us.
Later I asked for the instructions with three words, and I pointed at the box first:
From the recording, 41:05. There's no instructions for the agent yet. So just with three words, update agent instructions, it's going to start thinking again.
What did the instructions turn into?
When I clicked into the box afterward, I named three labels:
From the recording, 42:46. You'll now see, if you click into that box, my agent has operating rules. It has core jobs and it has a rhythm.
I also said all of it is editable, and that the easiest place to edit is the chat (42:57 to 43:03). The recording doesn't show the text under those labels. For a full set, the file on the session page is the one to read.
What does a full set of instructions say?
The file on the session page has eleven headings, and each one is a section you can write.
| Heading in the file | What the file says | | --- | --- | | Who you are | The role and the daily look at my calendar and inbox | | What you do | The daily brief plus drafted replies, in three parts | | What you read | Gmail, Calendar, research connectors, anything pasted in | | The daily brief | Today's calendar, emails that need a reply, open loops | | Prep before a call | Look the person up first, then give a one-page brief | | What you never do | The limits, as a list | | When I correct you | How a one-sentence correction is handled | | How you write drafts | Short, clear, in my voice | | Fallback | What to do when no connector is attached | | Where your answers go | Everything stays in the chat and in draft form | | What this version leaves out | What is not connected yet |
The file gives no times. It says "Every day" once, in the first section.
How do I write each section?
Take the sections in the file's order, and write yours in first person, as the file does.
- 01
Start with the role in a sentence.
The file opens with the role and what the AI agent does each day:
From the file on the session page. You are my Chief of Staff. Every day, you look at what's in front of me, my calendar and my inbox, or whatever I paste in instead, and you tell me what matters, draft what needs a reply, and propose where my schedule needs to move.
- 02
Say what comes back, and in what order.
The file says to give the brief "exactly this, in this order" and lists the three parts. The file on the session page says to name an overdue item plainly:
From the file on the session page. Name each one plainly. Don't soften it into a suggestion if it's actually overdue.
- 03
List what it reads, and what to do without a connector.
The file names Gmail, Google Calendar and three research connectors, and gives a fallback for pasted text.
- 04
Write the never list.
I give my AI agents guardrails: I tell them what they can do and what they can't do (41:26). The file puts the send rule in its never list:
From the file on the session page. You never send an email. Every reply is a draft, saved as a Gmail draft, never sent, until I say the words "send it" in that exact turn.
- 05
Say how a correction is taken.
The file asks the AI agent to keep the fix:
From the file on the session page. Keep the fix; don't quietly revert to the first version next time.
- 06
Say how drafts are written.
The file asks for short and clear writing in my voice, and for the other person's need to come first.
- 07
Close with what this version leaves out.
The file says what is not connected, and when to add it:
From the file on the session page. It does not connect to Slack, a CRM, or any other tool.
What goes wrong
The first is the one I raised in the session. The other two are rules in the file on the session page, not failures I raised.
- Instructions that update with nobody reading them. I left on the setting that lets the AI agent change its own instructions, and I said you want oversight into what it is updating, so put a review on the calendar (25:31).
How often to review is on its own page: How Often Should You Review an AI Agent's Instructions?. - A correction that doesn't carry over. The file asks the AI agent to keep a fix and not revert to its first version. - A draft that pretends a decision is already made. The file tells the AI agent to say so when a reply needs my decision first, and not to write the reply as if it were settled.
What the recording doesn't show: the text under operating rules, core jobs and rhythm. The file on the session page is a separate document, and I didn't read it aloud in the session.
Where this comes from
This article grows from the session I taught live with Women Applying AI on October 7, 2026. The quoted passages are my own words, taken verbatim from the recording or from the two files on the session page. The tables, steps and checklists around them were drafted by my team's AI agents from those words and my published pages, and checked against the recording before publishing.
Build Your First AI Agent in One Hour: the full session with the clips, the prompts and the transcript. The file is on the session page.
Related reading: What Are AI Operating Rules?, How Often Should You Review an AI Agent's Instructions?, What Does Edit Own Instructions Mean for an AI Agent?.
Frequently asked questions
Can the AI agent write its own instructions?
Yes. In the session I asked for them with three words, "update agent instructions", and the box went from empty to three labels: operating rules, core jobs and a rhythm.
Is there a section for what the AI agent must not do?
Yes, the file has one called What you never do. In the session I said guardrails are just as important as telling the AI agent what you want it to achieve (41:35).
What does the file say about corrections?
It asks the AI agent to update the next answer from a one-sentence correction and to keep the fix.
Build something that actually runs your workflow.
A focused, free 60-minute live session with Jacqueline. You build alongside her, on your own real task, and leave with an AI agent that is already running.