How to brief an agent so it doesn't waste your morning
Most people’s first agent disappoints them, and they blame the AI. Usually the AI did exactly what it was told. The instructions were just bad.
Briefing an agent is a skill, and it’s closer to managing a sharp new hire than to writing a search query. You wouldn’t tell a new analyst “keep an eye on the market” and expect anything useful on Monday. Same rules apply here. Here’s what separates instructions that work from instructions that produce three paragraphs of nothing.
Say what a good output looks like, not just what to watch
The most common mistake is describing the input and forgetting the output.
Weak: “Monitor our Google Analytics.”
Strong: “Each morning, check yesterday’s traffic against the same weekday last week. If anything moved more than 15%, tell me what and your best guess at why. If nothing moved, just say ‘nothing notable.’ One short paragraph, no charts.”
The second one tells the agent what to ignore, what threshold matters, and what shape the answer should take. “Nothing notable” is a feature. An agent that knows when to say nothing is worth ten that pad every brief to look busy.
Give it the thresholds you keep in your head
You already have rules about what deserves your attention. You’ve just never written them down, because for a human they’re instinct. An agent has no instinct. It has your instructions.
“Flag a thread only if it’s from a customer, mentions a deadline, or is the third message with no reply from us. Skip newsletters, calendar invites, and anything from the team channel.”
That one sentence does more than any model upgrade. It encodes the judgment you’d apply by hand, so the agent applies it for you at 6am instead of surfacing everything and making you triage anyway.
Tell it who it’s talking to
The same finding gets written ten different ways depending on the reader. Spell out who that is.
“Write it for me, not for a report. Assume I know the product and the team. Skip the throat-clearing. If you’re not sure something matters, include it but say so.”
Tone instructions feel optional. They’re not. They’re the difference between a brief you actually read and one you skim and close.
Be specific about cadence, then leave it alone
An agent that runs every fifteen minutes isn’t more helpful. It’s a notification you’ll mute by Thursday. Decide when you actually act on the information and run it then.
A market scan that you read with coffee should run at 6am, once. A “tell me the second a key account replies” agent runs more often, because that one is time-sensitive and you’d act on it immediately. Match the schedule to when the information changes your behavior, not to how often it could technically run.
The test for a good brief
Read your instructions and ask: could a competent stranger follow these and produce something I’d want? If the answer is “only if they already knew what I meant,” the brief isn’t done. Add the part you were keeping in your head.
The agents people love aren’t the ones with the best model behind them. They’re the ones whose owner spent ten extra minutes writing down what “useful” actually means. That ten minutes pays itself back every single morning.