How to Brief an AI Agent Properly

Three things you never have to say to a chatbot

You already know how to ask AI for things. Be specific, give context, say what good looks like. That advice is covered thoroughly elsewhere and it all still applies.

This lesson is about the three things you have to add the moment the AI can actually do something. None of them are about writing better instructions. They are about what happens when nobody is watching.

A finish line. What has to be true for this to be over.

Boundaries. What it must not touch, whatever else happens.

A stopping rule. The moment it should stop and come back to you.

A chatbot needs none of these, because a chatbot stops after one answer and touches nothing. An agent that lacks them will keep going, using its own judgement about when enough is enough. Its judgement is not bad. It is just not yours.

The finish line

Vague endings are the most common briefing failure. Help me with the newsletter has no end. An agent handed that instruction will do something reasonable and stop somewhere arbitrary, and you will not know whether it stopped because it finished or because it ran out of ideas.

A finish line is a description of the world after the work, not a description of the work.

Vague endings and their finished versions

VAGUE                        FINISHED
help with the newsletter  -> a draft in the shared folder, three
sections, under 400 words, subject
line suggested, not sent
look into the complaints  -> a document listing every complaint
from March grouped by theme, with
a count per theme
tidy the project files    -> a proposed folder structure written
into a plan document, no files moved
sort out the schedule     -> a list of the 4 clashes you found
and one suggested fix for each

A quick test for any finish line: could a colleague look at the result and say yes, that is done, without asking you what you meant. If not, the agent cannot tell either.

Boundaries

Boundaries are not politeness. They are the instruction that stops a reasonable decision from becoming an expensive one. And they have to be explicit, because an agent does not share your instincts about what is obviously off limits.

The useful trick is to state boundaries as things rather than as principles. Be careful with customer data is a principle, and it means nothing to an agent. Do not open anything in the Legal folder is a thing, and it works.

Boundary lines worth keeping in a note

Do not send anything. Save drafts only.
Do not delete or move any file. Copies are fine, originals stay.
Only look inside [this folder / this label / this date range].
Do not open anything marked confidential or personal.
Do not contact anyone outside the company.
Do not change any record. Report what you would change.
Do not spend anything or enter payment details anywhere.
Do not post publicly under any circumstances.
If you need a password or a login, stop and tell me.

In plain English

Brief:: The full instruction you hand an agent: the goal, the limits and the stopping rule together.
Finish line:: A description of what must be true for the job to be over, checkable by someone else.
Boundary:: A named thing the agent must not touch, stated as a specific place or action rather than a principle.
Stopping rule:: The condition that makes the agent pause and come back to you instead of deciding for itself.
Scope:: How much of your world this particular job is allowed to involve.

The stopping rule

This is the one people leave out, and it is the one that saves you.

Left alone, an agent handles surprises by making a decision. That is what it is for. The stopping rule tells it which surprises are above its pay grade. Without one, the agent that finds an unexpected invoice for a large amount will process it exactly as calmly as it processed the small ones.

Stopping rules you can lift straight into a brief

Stop and ask me if:

anything would be sent, published, paid or deleted
you find something that does not match what I described
a number is more than [X] or a date is outside [range]
the answer depends on a judgement about a person
you cannot find something you expected to find
doing the job properly needs access you do not have
you have been going for more than [N] steps without progress
When you stop, tell me what you were about to do and why you
paused. Do not pick the safest option and carry on quietly.

That last line matters more than it looks. An agent trying to be helpful will often route around a problem rather than raise it, and routing around problems is how you end up with a report that is missing March.

Checkpoint

Three additions turn a prompt into a brief: a finish line someone else could verify, boundaries stated as specific places and actions, and a stopping rule for the surprises that are not the agent's to resolve.

The whole thing in one shape

❌ Weak prompt

Prompt

Can you look at our supplier emails and follow up with anyone who has not sent their compliance document yet? Thanks.

Output

Reviewed supplier correspondence and sent follow-up requests to 14 suppliers. Two suppliers were sent escalation notices as they had been contacted previously.

It sent fourteen emails. Two of them were escalations, a word you never used, to companies you may be mid-negotiation with. No finish line, no boundary, no stopping rule, and every one of those emails is now in somebody's inbox.

✅ Good prompt

Prompt

Look only at the Suppliers label from the last 60 days. Find every supplier who has not sent a compliance document. For each one, save a polite chase email as a draft. Do not send anything and do not use the word escalation. Finish with a list of supplier name, date last contacted and what is missing. Stop and ask me if a supplier has already been chased twice, or if you are unsure whether a document counts.

Output

14 suppliers missing documents. 12 drafts saved, none sent. Paused on 2 suppliers already chased twice, listed at the top for your decision.

Same work, and the two cases that actually needed a human landed on your desk instead of in a customer inbox. The stopping rule did that, not the wording.

Write the boundaries before you write the task, not after. Boundaries added at the end of a long instruction get treated as an afterthought by you, which usually means one of them is missing, and the missing one is always the interesting one.

A template you can reuse

The four line agent brief

GOAL      [what should be true when you are done]
SCOPE     [exactly where to look, and nowhere else]
NEVER     [the actions that are off the table, listed]
STOP AND ASK IF  [the surprises that are mine to decide]
Then finish with: tell me what you could not access,
could not find, or had to guess.

Four lines. It takes about ninety seconds to write and it is the difference between an agent that saves you an afternoon and one that costs you a week.

📝 Quiz

Question 1 of 4

What does an agent brief need that an ordinary chatbot prompt does not?