How to write the first job you give an AI coworker

Write the first job you give an AI coworker like a brief for a new colleague. Name the result you want, why you need it, the tool it should use, what done looks like and what you will decide yourself. Start with a small read you can check, such as a list, before you schedule anything.

What makes a good first job for an AI coworker?

A good first job is one everyday task you know well, set as a small read you can check. A 90-day plan on Microsoft’s Signal blog, taken from the book Open to Work, tells readers to “Choose one everyday task you already do on autopilot”. It then asks them to note “what helped, what didn’t and what you’ll try next”.

The same plan sorts work into buckets. One is tasks AI can do alone, like routine reports, data entry and scheduling. Another is “Uniquely human tasks, like building trust, making ethical decisions and reading a room”. We keep those with people.

We suggest starting with the read behind a report, not with data entry. A list or a lookup changes nothing in your tools, so a miss is easy to fix. Our guide to using AI in Slack covers the day-to-day asks. This post is about how to write the brief.

What should the first message include?

The first message needs a clear goal, and often a little more. Microsoft’s Copilot prompt guide lists four parts: Goal, Context, Source and Expectations. It says “all that’s required is a clear goal”, and that you’ll often need more than a goal to get the results you want.

Anthropic’s prompting guide for developers asks you to think of its model as a “brilliant but new employee who lacks context on your norms and workflows”. In our view, an AI coworker starts the same way. It does not know which list you trust or why today matters.

The table adds a fifth part of our own: what you will decide yourself. Slack’s guide for agent builders says an agent taking an action with real-world effects, such as creating, sending or deleting content, should require explicit confirmation. Writing your decisions into the brief makes clear which steps stay with you. One line for each part is enough.

These examples are not about any one product.

How do you write it, step by step?

Write the brief in six short steps, then test it on a colleague.

1. Pick one small read with an answer you can check.
2. State the result in one sentence, such as “list the open requests”.
3. Add why you need it and who will use it.
4. Name the tool or source it should use.
5. Say what done looks like, for example a list posted in the thread.
6. Say what you will decide yourself.

Then run a test from Anthropic’s guide. Show the brief to a colleague with minimal context on the task and ask them to follow it. Anthropic says that if they’d be confused, its model will be too.

Example. Mei runs operations at Northwind and wants to hand a timesheet job to HiToby. Her first try reads “@HiToby sort out timesheets.” It does not say what to return, where to look or what she will decide.

Her rewrite reads “@HiToby list the unapproved timesheets in Deputy and post them in this thread with each person’s name. Payroll runs tomorrow, so I need to see who is still waiting. I will approve them in Deputy myself.”

Example reply: a list of names with the timesheets still waiting.

The rewrite has all five parts. The result is a list, the context is payroll, and the source is Deputy. What done looks like is names in the thread, and approval stays in Deputy, with Mei. Our guide to checking Deputy timesheets from Slack shows how HiToby reads Deputy.

How do you check the first result?

Check the first result against the tool itself or against what you already know, and ask what it left out. Microsoft’s Copilot prompt guide says to “Evaluate Copilot’s responses and cross-reference with trusted sources when needed”.

Slack’s guide for agent builders says that when an agent finishes a task, it should give “a clear summary of what happened, including what was skipped and why”. If a reply does not say what was left out, ask.

Expect to go round more than once. Microsoft’s Copilot prompt guide says “Expect some back-and-forth conversation to get the results you’re looking for”. When a result misses, change one part of the brief and ask again. If the answer came from the wrong place, fix the source line. If it came back in the wrong shape, fix the line on what done looks like. Changing one part at a time helps show which part was weak.

Slack’s guide for agent builders also warns that asking for confirmation on every action “creates fatigue and trains users to click through without reading”. A brief can follow the same idea. Save “check with me first” for steps that change something.

When should a job go on a schedule?

Put a job on a schedule once the one-off result comes back right. A schedule runs the job without you asking again, so a vague brief can miss again on later runs. Microsoft’s Copilot prompt guide also notes that “Using the same prompt multiple times can result in different responses.”

Our suggested order:

1. Run the job once by hand.
2. Fix the brief until the result matches what you expected.
3. Read the first few scheduled results with the same care as the first.
4. After that, keep reading them, even if you check less closely.

A report with the same shape each week suits a schedule. A job that needs judgement about people stays with a person, as the first section said. For the wider idea, see what an AI employee is.

Where does HiToby fit?

HiToby is an AI coworker that lives in Slack or Microsoft Teams, and you can hand it work with a brief like this one. You message HiToby in plain English, like a colleague, or mention it in a thread. It breaks the request into steps, uses the tools you connected and posts the finished result in the thread.

You can give it a schedule and it runs a job, such as a morning brief or a weekly report, without being asked again. HiToby strives to ask before anything goes out under your name, and before an external action it shows an Approve or Reject step.

Setup takes about two minutes. You sign in with Slack or Teams, choose which channels it can see, connect a tool if you want to, and give it something to do.

You can use this method with other AI coworkers too. To try it with HiToby, book a demo or look at pricing.

Slack is a trademark of Slack Technologies, LLC. Deputy is a trademark of its owner. Microsoft Copilot and Microsoft Teams are trademarks of Microsoft Corporation.

Questions

What should I ask an AI coworker to do first?

A small read you can check, such as a list or a lookup in a tool you already use. Write the brief with the result, why you need it, the source, what done looks like and what you will decide.

How long should the instruction be?

Short, about one line for each part. Microsoft’s Copilot prompt guide says “all that’s required is a clear goal”, and that you’ll often need more to get the results you want. Give the result, why you need it, the source, what done looks like and what you will decide, then change one part if the first try misses.

Should an AI coworker ask before it acts?

Yes, before actions with real-world effects. Slack’s guide for agent builders says an agent should require explicit confirmation for those actions, and warns against asking on every action. HiToby strives to ask before anything goes out under your name.

When can I give an AI coworker a recurring job?

Once the one-off result is right. With HiToby you can give it a schedule and it runs a job, such as a morning brief or a weekly report, without being asked again.