The Five-Field Brief That Stops AI Agents Running Wild

Give the agent a job, not an ambition

An agent handed “improve our email marketing” will do something, confidently, today. Whether that thing is worth doing is a coin flip, because the instruction contains no job, only an ambition. The single biggest gap between agents that quietly help and agents that create cleanup work is task definition. Write a brief the way you would for a capable new employee: small scope, fixed tools, named limits, a way to ask for help. Here is the structure that takes five minutes and pays for itself the first week.

The brief, in five fields

GOAL:     one sentence, a verb and a finish condition
SCOPE:    which accounts, pages, or inboxes; nothing else
TOOLS:    named list, nothing outside it
GUARD:    what it may never do without asking you
CHECK-IN: when you will next read its work

A real brief filled in, for morning shipping checks:

GOAL:     Each morning, read new support emails and tag each
          one: order-status, return, product-question, other.
SCOPE:    The support inbox only. No other account.
TOOLS:    Email read + label. No send button, ever.
GUARD:    Never reply to a customer, never delete anything,
          never change a label that a human wrote.
CHECK-IN: I review yesterday's labels with coffee each morning.

Narrow, inspectable, and safely failed: if the agent does the wrong thing all morning, the worst harm is a label you undo in a click, because you designed the blast radius before the first run.

Why “one job” beats “one ambition”

The mistake most first-robot owners make is the same one new managers make: a sprawling instruction feels like generosity and behaves like blame. “Sort the inbox, answer shipping times, flag angry customers, and summarise the rest” is four jobs, each needing its own brief. Ship one, watch it work twice, then ship the next. The agent does not tire of small jobs the way an intern would, and you learn what it is actually good at, one task at a time.

Guard first, capability second

Write the guard line before you pick the tools, not after the first surprise. Two rules cover most of the risk:

GUARD: Never spend money, send email, or contact a customer.
GUARD: Never delete. Move and label only, a weekly review can undo any of it.

Reversible actions only, until the check-ins build enough trust to widen the scope. Every access you grant should have a reason and an owner’s name, like any other account in the business.

Schedule the check-in before the agent starts

The review is a calendar event, not a feeling: put fifteen minutes in the diary for the first week, then weekly. Read the log, undo anything that drifted, and tighten one rule if the log shows a pattern you did not intend. This is how agents get better, with small instructions after real runs, not with a longer first brief.

The one-sentence test for any brief

If you cannot read the goal to a stranger on a train in one sentence and have them know when it is finished, the brief is not ready. “Tag each support email by type, every morning, before 9am” passes. “Help with the inbox” does not. Write the five fields, run it once, and you will have an agent that helps the way a good hire does: quietly, within scope, and worth a coffee.

Want more like this? Join MOKU Club for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.