How to Write Tasks Your AI Employee Actually Nails

Key Takeaways

• A good task reads like a brief you would give a sharp new hire, not a search query. Say what you want, where the inputs are, and what the finished thing looks like.
• Name the source and the destination. "Pull from the #sales channel and the HubSpot pipeline, post the result here" beats "check sales" every time.
• Describe the output shape. Five bullets, a table, a draft reply, a short Slack post. If you do not say, you get the AI employee's guess.
• Set the boundary in the task. Add "show me before sending" or "read-only" and you keep control without slowing anything down.
• Vague in, vague out. Most disappointing results are not a capability problem, they are a briefing problem, and briefing is a two-minute skill to learn.
• You can iterate in the thread. Treat the first result as a draft, reply with one correction, and it adjusts. Viktor remembers your fixes so you brief it less over time.

The task came back wrong, and it was not the tool's fault

You typed "summarize the sales calls" and got back something technically correct and completely useless: a wall of text when you wanted five bullets, this week's calls when you meant last week's, no mention of the one deal you actually cared about. It is tempting to conclude the AI employee is not smart enough. Usually the opposite is true. It did exactly what you asked. The problem is that "summarize the sales calls" is not a task, it is a hint.

The Slack 2024 Workforce Index found that desk workers spend about a third of their day on tasks they consider low-value. Handing those tasks to an AI employee is the whole point, but only if you brief it well enough to get back something you can use. So here is the format that turns a vague ask into a result you would have been happy to get from a person.

What makes a task "good" for an AI employee?

Answer first: a good task states the goal, points to the inputs, and describes the output. Miss any one of those and the AI employee has to guess, and its guess is where your disappointment comes from.

Think about how you would brief a capable new hire. You would not say "handle marketing." You would say "take last week's Google Ads spend, compare it to the week before, and put the three biggest changes in a short Slack post for the team." That sentence has all three ingredients, and a new hire could run with it. Your AI employee needs the same three things:

• Goal: what you actually want to end up with.
• Inputs: where the information lives (which channel, which tool, which time range).
• Output shape: the format and length of the finished thing.

Everything below is just how to make those three explicit without turning every request into an essay.

Name the source and the destination

The single biggest upgrade to any task is telling your AI employee exactly where to look and where to put the result. "Check sales" could mean a Slack channel, a HubSpot pipeline, a Notion doc, or all three. Naming the source removes the guesswork.

Compare these two:

@Viktor how are we doing on sales?

@Viktor look at the #sales channel from this week and our new-business pipeline in HubSpot. Post a 5-bullet summary here with total closed, biggest open deal, and anything stuck more than 5 days.

The second one names the source (the channel plus the HubSpot pipeline), the time range (this week), the destination (here), and the shape (5 bullets with three specific things). It cannot really miss. The first one is a coin flip.

Describe the output shape

Answer first: if you do not say what the finished thing should look like, you are letting the AI employee choose, and it will often choose "a paragraph." Name the format and you get the format.

Output shapes worth asking for by name:

• Bullets for scannable summaries. "Five bullets, one line each."
• A table for comparisons. "A table with columns for owner, deal, and days stuck."
• A draft for anything that will be sent. "Draft a reply I can review, do not send."
• A short Slack post for team updates. "One short paragraph the team can read in ten seconds."

The output shape is also where you control length. "A short summary" and "a thorough write-up" produce very different results from the same inputs, so say which one you want.

Put the boundary inside the task

A task is also where you set the rules of engagement, and this is what keeps delegation from feeling risky. Two short phrases do most of the work:

• "Show me before sending." For anything that leaves your workspace, like an email or a message to a customer, this keeps a human checkpoint in the loop while still doing all the heavy lifting.
• "Read-only" or "do not change anything." When you just want information, say so, and the AI employee will look without touching.

This connects directly to how you set up access in the first place, which we cover in How to Control What Your AI Employee Can Access. Access settings are your standing policy. A phrase in the task is your per-task override, and using both is how careful teams move fast without losing control.

Bad task vs good task, side by side

The gap between a fumbled result and a great one is almost always in the wording. Same tools, same AI employee, different brief.

None of the good versions are long. They are just specific about goal, inputs, and shape.

What if the first result is not quite right?

Answer first: reply in the same thread with one correction. You do not have to rewrite the whole task, you adjust it like you would with a person.

If the summary is too long, "make it half as long." If it pulled the wrong range, "just last week." If it missed something, "also include the deals that slipped." The AI employee keeps the context of the thread, so each reply refines the result instead of starting over. This is where working with an AI employee feels different from a search box: it is a back-and-forth, not a one-shot query.

And the corrections compound. When your AI employee has real memory, the fix you gave today ("we always want revenue net of refunds") becomes the default tomorrow, so you brief it less over time. We wrote about how that memory works in How to Make Your AI Employee Actually Remember. The practical effect: your tasks get shorter as the AI employee learns your preferences, because you stop having to repeat them.

Turn a task you repeat into a saved routine

If you find yourself writing the same task every Monday, that is a signal to turn it into a standing routine rather than retyping it. A well-briefed recurring task ("every Monday at 9, post last week's Google Ads and Meta Ads changes to #growth in five bullets") runs on its own and lands in the right place, in the right shape, without you rewriting the brief each time. The same three ingredients apply, you just set them once. For ideas on what is worth making recurring, see Recurring Tasks Your AI Coworker Should Own.

Frequently Asked Questions

Why does my AI employee give me the wrong thing?

Almost always because the task was underspecified. If you do not name the source, the time range, and the output shape, the AI employee fills the gaps with its best guess. Adding those three things fixes the large majority of disappointing results.

How detailed does a task need to be?

Detailed enough to state the goal, point to the inputs, and describe the output, which is usually one or two sentences. You are writing a brief for a capable colleague, not a program, so plain language is fine as long as it is specific.

Can I correct a result without starting over?

Yes. Reply in the same thread with the change you want, like "make it shorter" or "just last week," and the AI employee refines its previous answer while keeping the context. It is a conversation, not a single query.

How do I stop it from sending something before I have seen it?

Put the boundary in the task: add "show me before sending" or "draft only, do not send." For a standing rule, you can also keep a tool connected in read-only mode so it cannot send or change anything at all.

Do my tasks get easier over time?

They do when your AI employee has memory. Corrections you make once, like a preferred format or a definition of a metric, become defaults it applies going forward, so you can brief it more briefly as it learns how you work.

What is a good first task to practice on?

Something small and easy to check, like "summarize the last 20 messages in this channel in five bullets." Verifying a small result teaches you how the AI employee interprets a brief before you hand it anything important.

Brief it well and it works like a great hire

Getting real work out of an AI employee is less about clever wording and more about a clear brief: say what you want, point to where the inputs are, and describe the finished thing. Add a boundary when the task touches the outside world, correct it in the thread when needed, and turn the repeats into routines. Do that and the tasks eating a third of your day stop being your problem and start being handled where your team already works.

Add Viktor to your workspace and hand it a first task

Related reading:

• How to Control What Your AI Employee Can Access
• How to Make Your AI Employee Actually Remember
• Recurring Tasks Your AI Coworker Should Own