Stop Asking Cowork to Do Everything in One Sentence

Open Cowork, type "summarize my unread emails", and hit enter.

You will get a summary, but it will be the wrong summary. It'll include the newsletter you never read, the calendar invite you already accepted, and the marketing email from a vendor you've been ignoring for the last year. The two messages from clients that need a response today will be in there somewhere, buried.

You'll close the tab and decide AI isn't quite there yet.

But the AI is fine; your brief is the problem.

#The Google-search habit

People naturally prompt Cowork the way they search Google: a few keywords, hope for the best, refine if it sucks. This is fine when your goal is genuinely simple, like a quick definition or a one-step lookup. It doesn't work when your goal involves judgment, which is almost every task worth automating.

The reason is structural. A search engine ranks pages while Cowork makes decisions. Decisions need context; five keywords don't carry enough context to make a decision you'd actually trust.

#Brief it like a new hire

The highest leverage fix for bad Cowork output is to stop prompting and start briefing.

Suppose you've hired the perfect new assistant. On day one, they've never met you, they don't know your colleagues, they don't know what "the Q3 thing" means, and they have never seen your inbox. They're smart and capable, but they need a brief.

You'd never send this assistant a Slack that says "summarize my emails." You'd say something more like: "Look at my unread Gmail from the last 24 hours. I care about anything where I owe someone a reply, especially from clients. Skip newsletters and marketing. Give me a list including the sender, a one-line summary, and what kind of response is needed. Keep it under 15 items."

That's an effective brief, and Cowork can do the rest.

#The four parts of a good brief

A useful brief has four parts.

The goal. What does done look like? This is not the task; it is the outcome. "Summarize my emails" is a task. "Help me decide which emails I need to handle before noon" is an outcome. The outcome tells Cowork which details matter.

The constraints. What should be left out? What should be prioritized? "Skip newsletters" carries more information than three sentences of positive instruction, because it tells Cowork what you already know how to filter on your own.

An example of done. Describe the shape of the output. If you don't do this, Cowork will guess, and its guesses tend to be verbose and generic.

Where the inputs are. Cowork can't read your mind, your calendar, or your inbox unless you point it at them. "Look at my unread Gmail" is a pointer. "Read the doc called 'Launch Plan' in the marketing folder" is a pointer. Without pointers, Cowork has to ask, or it hallucinates.

#Before and after

Here's the same request, briefed the wrong way and the right way.

Before: "Summarize my unread emails."

After: "Look at my unread Gmail from the last 24 hours. Pull out anything where I owe someone a reply, anything from a client, and anything from my manager. Skip newsletters, calendar invites, and marketing. Format as a numbered list with sender, a one-line summary, and what kind of reply seems needed. Keep it under 15 items; if there are more, prioritize the highest-stakes ones."

The second version is longer because it carries a goal (decide what to handle), constraints (what to skip, how many), a format (numbered list with three fields), and a pointer (unread Gmail, last 24 hours). All four parts.

You can run that prompt and trust the result.

#When the long brief is overkill

The four-part brief is not the rule for every prompt. For genuinely simple, single-step asks, the Google-search habit is fine.

The brief is necessary when:

• The task involves judgment (what's important, what's relevant)
• Multiple inputs are in play (calendar plus inbox plus a doc)
• The output is something you'll act on (a draft, a list, a decision)
• You'll want to run the same kind of request again later

If at least one of these is true, I put effort into a brief.

#The compounding move

A good brief is reusable.

Pick one task you do every week, and write a four-part brief for it. Save it as a markdown file in your Cowork workspace. Next week, paste it in. The week after, tweak it based on what you've seen during the last couple weeks. By the third week, you have a prompt that works better than the version you'd write from scratch, and you'll never write it from scratch again.

Lots of people stay stuck in the Google-search habit because they don't think of prompts as artifacts. Prompts can be written, saved, edited, reused, and improved like any other piece of work.

#What to do this week

Pick a task you do frequently, set a time for 5 minutes, and write a brief with four parts: goal, constraints, example of done, where the inputs are. Save it as a .md file somewhere you'll find it again.

The next time that task comes up, you're ready to brief Cowork like a colleague instead of a search engine.