How to Write a Task Brief Your AI Agent Can Actually Use
The brief matters more than the model. Give your AI agent the outcome, the constraints, and the shape of a good result, and watch the quality jump.
By Professor Muse4 min read
Most disappointing agent results come from thin briefs. (For the general version of this skill, see how to get genuinely useful results from your AI assistant.) The model is usually fine. The agent only knows what you tell it, and unlike a human assistant, it cannot read the room. Five extra sentences up front routinely save five rounds of corrections.
The short version
- State the outcome: what “done” looks like.
- Add the context a smart colleague would need: constraints, budget, deadlines, preferences.
- Describe the shape of the result: format, length, and what to do with edge cases.
- Iterate like an editor. Specific corrections beat re-sending the same vague brief.
Brief the Outcome
Bad briefs describe steps. Good briefs describe the finish line and let the agent choose the route. “Check three airlines for Friday flights” is a step. “Book me the cheapest nonstop Friday flight to Denver that lands before 6pm, under $300” is an outcome. The agent can compare sources, weigh tradeoffs, and come back with a decision for you to approve. The step version just generates busywork you then have to assemble.
When you catch yourself writing steps, stop and ask what you would accept as a finished result. That answer is your brief. The steps were never the point; they were your guess at how to get there, and the agent’s guess is often better informed.
The Four Parts of a Working Brief
1. The outcome. One sentence saying what “done” means. “A booked table for two.” “An inbox with everything either answered, delegated, or deleted.” If you cannot state the outcome, you are not ready to delegate.
2. The context. What a capable colleague would need: who it is for, what you have already tried, the tools and accounts involved. “Use my Gmail and Google Calendar. I already checked OpenTable and everything downtown is full.”
3. The constraints. Budget, deadline, and hard limits. “Under $150 total. Book for next Friday. Do not email anyone without my approval.” Constraints are where agents earn trust, because they turn judgment calls into rules.
4. The shape of the result. Format, length, and how to handle the unclear bits. “Reply with a shortlist of three, each with price and one reason you picked it. If nothing fits the budget, tell me the closest option instead of guessing.”
A Before and After
Before
"Plan my trip to Austin."
After
"Plan a 3-day Austin trip, March 12 to 14, two people, hotel under $200 a night near downtown. Find one live-music night and one barbecue spot locals actually like. Send me an itinerary with prices; do not book anything until I approve."
The first brief invites twenty questions. The second gives the agent everything it needs to work for an hour and return something decidable.
Iterate Like an Editor
First attempts still miss. Correct them the way an editor corrects a draft: point at the specific thing. “The hotel is too far from downtown; keep it within a mile of Congress Avenue” beats “try again.” Each correction teaches the agent your preferences, and many agents remember them across tasks.
Answer Its Questions Well
A good agent asks before it guesses. When Muse comes back with three clarifying questions, treat them as cheap: answering takes a minute, while a wrong hour-long task costs far more. Answer each one specifically instead of waving it through with “use your best judgment.” The exception is low-stakes exploration, where “surprise me” is a legitimate instruction. Reserve it for tasks where any outcome is acceptable.
Save the Briefs That Work
One-off briefs are only half the value. The real payoff comes from briefs you reuse: the weekly report, the trip template, the inbox rules. Write these once, carefully, and save them where you can paste them in. Better, ask the agent to remember the durable parts as standing instructions: your budget ranges, your airline preferences, the format you like options in. A standing instruction is a brief you never have to write again, and it compounds. After a month, your “brief” for a familiar task is one sentence, because the other twenty live in memory. When a brief produces a great result, keep it, and keep refining the library.
A good brief is a skill, and it transfers. The same habits improve results from Muse, ChatGPT, Grok, and whatever comes next, because every agent is only as good as its instructions. Start with your next real task. Write the brief, watch what comes back, and correct one thing. That loop, repeated, is the whole skill.
