How to brief an AI (or a person) so the work comes back right.
Most bad AI output starts with a bad handoff. Five things every brief needs, and three it shouldn't have.
I've watched a lot of AI work go wrong, and when I trace it back, the model is rarely the problem. The brief is. The worker didn't know what done looked like, or worked on the wrong copy of a file, or repeated something that had already failed, or followed an instruction that had quietly been replaced.
The funny thing is that none of this is new. It's exactly how work goes wrong between people. Agents just make it happen faster and more often, so the lessons get impossible to ignore. Here's what I've learned to put in a brief, and what to leave out.
1. The goal, in one sentence
Not the task, the goal. "Write a landing page" is a task. "Get small studio owners to join the waitlist" is a goal. The goal lets the worker make good calls on all the little decisions you didn't think to specify.
If you can't write the goal in one sentence, the worker definitely can't infer it.
2. What done looks like
This is the one people skip most, and it causes the most pain. Without it, the worker stops when it feels done, which is usually too early. Say what the finished thing is, where it should end up and how you'll judge it. "A one-page proposal, in the proposals folder, that a client could sign without a follow-up call" is something a worker can aim at and check against.
3. What it's allowed to do
Can it send the email or only draft it? Spend money? Change things other people own? Say so up front. A worker that doesn't know its limits either asks you about everything or does something you didn't want.
With agents, the important limits shouldn't live only in the brief. They should be enforced by the system too, so a misread sentence can't turn into a sent email. But saying them in the brief still helps the worker plan.
4. What's already been tried
If something failed before, say what and why. "We tried a free trial and nobody converted" saves the next worker from rediscovering it. This is the most commonly lost information in any handoff, because successful work gets kept and failed work gets binned.
Be specific. "The layout felt off" is hard to use. "On a phone, the image covered the sign-up button" is something the next attempt can actually fix.
5. Exactly where things live
Name the actual file, folder, document or page. Not "the deck", but which deck, and which version. A surprising amount of wasted work comes from a worker editing a copy, or a folder that doesn't exist, or last month's version of a doc.
For agents this matters even more, because they'll happily invent a plausible-looking location if you're vague. A good system treats file locations as exact facts and refuses made-up ones.
And three things to leave out
More isn't better. These three make a brief worse:
- The whole history. Forwarding the full conversation feels thorough, but it buries the two facts that matter under everything that didn't. Pick what's relevant.
- Guesses dressed up as facts. If you're not sure, say so. A worker treats everything in a brief as true, and a confident guess becomes a wrong assumption.
- Instructions that no longer apply. Old instructions sitting next to new ones get followed about half the time. Delete them, or mark them clearly as replaced.
A template you can steal
Goal: one sentence on why this matters.
Done looks like: the finished thing, where it lives, how it'll be judged.
You may: what it can do alone, and what needs a yes first.
Already tried: what failed and why, specifically.
Where things are: exact files, documents and links.
Not sure about: anything you're guessing at.
It works for agents, contractors and colleagues alike. Which, honestly, is the whole point: briefing an AI well is just management, done carefully.