## Key Takeaways

- **Most disappointing AI output traces back to the brief, not the AI.** The same task, briefed the way you would hand it to a capable new hire, routinely comes back at a different quality level.
- **A good brief has five parts:** the outcome, the sources, the format, the audience, and the edge cases. You rarely need more than five sentences to cover them.
- **Name your tools and places.** "Check the numbers" is a guess. "Pull closed-won deals from HubSpot for last week" is a task.
- **Delegate outcomes, not keystrokes.** You are not writing code. Say what done looks like and let your AI employee choose the route.
- **The first draft of a recurring task is a conversation.** Run it once, correct what is off, and ask him to remember the corrections. The second run is where the payoff starts.
- **Steal from your own best delegation.** The brief you would give a trusted colleague on their third week is almost exactly the brief that works.

## The quiet reason AI employees underperform

A founder asks their AI employee to "look at our marketing performance" and gets back a generic overview they skim and discard. The conclusion feels obvious: not ready for real work. Then someone on the same team asks for "Meta and Google Ads spend versus last month, broken out by campaign, flag anything whose cost per lead moved more than 20 percent, table format, for Monday's growth meeting" and gets back exactly that, in minutes.

Same AI employee. Same tools. Same data. The difference was the brief, and this is the least discussed skill in working with an AI employee: delegation quality transfers. People who brief humans well get good output almost immediately. People who have never had to delegate treat the AI like a search box and conclude it is one.

This post is the playbook for the first group's habit, written down for everyone else.

## The five parts of a brief that works

You do not need prompt engineering. You need the same five things a capable colleague would ask you if the brief were vague, answered in advance.

**1. The outcome.** What exists when this is done? A table, a document, a sent draft, an updated sheet? "Research our competitors" has no finish line. "A one-page comparison of how our top three competitors price onboarding, ending with the two things we should copy" does.

**2. The sources.** Name the tools, channels, or documents to draw from. "Pull last week's tickets from the support inbox" beats "look at support volume", because naming the source removes the largest guess. If the source is a conversation, say which channel.

**3. The format.** Table, bullets, PDF, spreadsheet, a message in the channel, a draft email. Format is cheap for you to specify and expensive to guess wrong.

**4. The audience.** Work for the board looks different from work for the intern channel. One clause is enough: "for a customer", "for our Monday standup", "for someone who has never seen our metrics".

**5. The edge cases.** The one or two judgment calls you can predict. "If a deal has no close date, put it in a separate list rather than guessing." This is the part that separates a brief from a wish, and it is where most rework hides.

Here is all five in one real-shaped example:

```prompt
@Viktor pull all invoices from the finance inbox from June, put them in
a spreadsheet with vendor, amount, due date, and status. If an email has
an attachment that isn't an invoice, skip it and list it separately at
the bottom. Format for our bookkeeper, she'll work directly in the
sheet. Post it here when done.
```

Outcome, source, format, audience, edge case. Five sentences, no special syntax.

## Delegate outcomes, not keystrokes

There is a failure mode on the opposite end: the brief that micromanages. Ten numbered steps, each specifying a click. It feels rigorous, and it usually produces worse results, because a step-by-step script removes the AI employee's ability to route around surprises. If step 4 assumes a report exists and it does not, a script breaks where a goal would adapt.

The working rule: **specify what done looks like as precisely as you can, and the route as loosely as you can.** "Reconcile the sheet against the CRM and tell me what doesn't match" leaves the method open. If you find yourself writing step 7, ask whether steps 1 through 6 are actually constraints or just your habit.

The exception is genuine constraints: things that must or must not happen. "Don't email anyone, drafts only." "Use the numbers from the finance sheet, not the dashboard." Constraints belong in every brief they apply to; choreography belongs in none.

## The first run is a conversation, not a verdict

The most expensive mistake teams make is treating the first output as a final exam. A capable new hire's first draft of your weekly report is not their best work either; it becomes their best work because you correct it once and they remember.

Work the same loop with your AI employee. Run the task, then correct specifically: "the table should be sorted by amount, and next time exclude the internal test transactions." Then make it stick: ask him to remember those corrections for this task going forward. That turns a one-off brief into a standing standard, and it is how a five-sentence request quietly becomes a process that runs the same way every week: the corrections persist, so the standard compounds instead of resetting with every ask.

By the third run of a recurring task, the brief has usually shrunk to one line, because the standard carries the rest.

## Before and after: three real briefs

| The wish | The brief |
| --- | --- |
| "Look into our churn" | "From the CRM, list customers who cancelled in the last 90 days with plan and tenure. Group by cancellation reason where we have one. End with the three patterns you see. Doc format, for our retention discussion Thursday." |
| "Can you handle the meeting notes?" | "After my external calls, post a summary in #sales-notes: attendees, what they asked for, what we promised, and next steps with owners. If a call has no clear next step, flag that instead of inventing one." |
| "Research this company" | "One-page brief on Acme Corp before my 2pm: what they do, headcount trend, recent news, and who their buyer likely is. End with three questions worth asking them. I'll read it on my phone." |

Read the left column out loud and you can hear the problem: you would not hand those to a person either.

## Frequently Asked Questions

### Do I need to learn prompt engineering to work with an AI employee?

No. The five-part brief is management, not engineering: outcome, sources, format, audience, edge cases. If you can delegate to a person, you already have the skill, and if you cannot, practicing on an AI employee is the cheapest training available.

### How long should a brief be?

Usually three to six sentences. Shorter than that and you are making him guess; a screen-long brief usually means you are scripting keystrokes instead of describing an outcome. Recurring tasks shrink over time as corrections become remembered standards.

### What if I don't know what format I want?

Say so and ask for a proposal: "suggest a format and show me a sample with three rows before doing the whole thing." A thirty-second preview is much cheaper than a full deliverable in the wrong shape.

### Why did I get a generic answer to a specific question?

Almost always a missing source or audience. "How is marketing doing" invites a generic answer; naming the ad accounts, the time window, and who the answer is for pins it to your data. If a named tool was not used, check that its integration is connected.

### Should I correct mistakes or just redo the brief?

Correct them, specifically and once. "The revenue column should come from Stripe, not the sheet" fixes this run and, if you ask him to remember it, every future run. Rewriting the brief from scratch each time throws away the accumulated standard.

### How do I brief a task I only half understand myself?

Delegate the understanding first. "Ask me the questions you need answered before doing X" turns the AI employee into the colleague who scopes the work with you. The brief that follows those questions is usually the best one you will write.

## The brief is the skill

Teams keep looking for the setting, the model, or the trick that makes AI output good. For an AI employee doing real delegated work, the highest-leverage variable was never hidden: it is the same delegation quality that separates good managers from bad ones. Write the outcome, name the sources, fix the format, know the audience, call the edge cases. Then correct once and let the standard stick.

[Add Viktor to your workspace and put a good brief to work](https://viktor.com/?utm_source=blog&utm_medium=cta&utm_campaign=how-to-write-a-brief-your-ai-employee-can-run-with)

Related reading:

- [What Is an AI Coworker?](/blog/what-is-an-ai-coworker)
- [Replace Meetings With Async Reports](/blog/replace-meetings-with-async-reports)
- [Replace Weekly Reporting With AI](/blog/replace-weekly-reporting-with-ai)