Hand a whole task to AI and stay in control

What you will learn

• How projects, memory and AI agents let you hand over a whole task, not one paragraph at a time
• How to write a brief: goal, inputs, what done looks like, and limits
• How to stay in control with a plan you approve, check-in points and a final review

The idea

Many repeated tasks are bigger than one prompt, such as a weekly report. Three features help you hand over the whole task. ChatGPT, Claude, Microsoft Copilot, Gemini and Perplexity are examples of products with some of them. The tool pages show which product has which, as far as we have confirmed it, and on which plans.

A project, or workspace, keeps the files and standing instructions for one kind of work in one place. You upload last month's report and write "same structure, under one page" once, and every chat in the project starts from there. You stop repeating yourself.

Memory is what a tool keeps about you between chats, such as your role or preferred style. Often you can see what it has saved, delete items or turn memory off. Check it now and then: an old or wrong detail can quietly shape new answers.

An AI agent works towards a goal over several steps, deciding what to do next and using tools such as search, files or other apps. Some can browse websites, fill in forms or work in your email and documents. That is also where things go wrong: agents can misread instructions or be tricked by hidden ones. Give an agent access only to what the task needs.

Brief the tool as you would a new colleague. Give the goal, the inputs, what done looks like, and the limits, such as "do not send, delete, buy or share anything". Ask for a plan first and approve it before the work starts. Set check-in points where it stops and shows you what it has. Then review everything before anything leaves your hands.

Example

Jonas, an office manager, compares three cleaning companies every year. He creates a project with last year's comparison table and writes:

Goal: compare three office cleaning companies so my manager can choose one.
Inputs: their websites only: [link 1], [link 2], [link 3]. Use the
columns from last year's table in this project.
Done means: one table with services, published prices, contract length
and contact details, and a source link for every cell. Write "not
found" where a website does not say.
Limits: read only. Do not fill in forms, sign up, send, buy or share anything.
First show me your plan and wait for my OK.
Stop after the first company and show me the result.

Jonas adds a step to the plan: check which area each company covers. At the first check-in he sees a price from a review site, not the company's own page, and asks for official pages only. Before he sends the table to his manager, he opens every source link for prices and contract terms.

Practice now

1. Choose one small task you repeat, such as a weekly report or a comparison of three suppliers from their websites.
2. If your tool has projects, create one for this task with a past example and your standing instructions. Leave out files your company or school does not allow.
3. Open your memory settings and delete anything wrong or unwanted.
4. Fill in the template below, including the limits.
5. Read the plan, change it if needed, then approve it.
6. Review at each check-in, then check the result against your "done means" line.

At the end you should have a finished task you reviewed yourself, and a brief you can reuse. Next, read Before you use AI at work and the guide to AI at work.

Copy this template

Task: [the whole task in one sentence]
Goal: [what the result is for and who will use it]
Inputs: [files, links or notes to use, and nothing else]
Done means: [format, length, and what each point needs, such as a source link]
Limits: Do not send, delete, buy or share anything. Also do not [other limits].
If information is missing, write "not found" instead of guessing.

Before you start, show me a short plan and wait for my approval.
Check-in: stop after [step] and show me what you have.
At the end, list what you could not do and anything you are unsure about.

Asking AI to flag uncertainty does not make its answer correct. Keep checking the parts that matter.

Check the output

• Does the result meet every point in your "done means" line?
• Did it use only your inputs and stay inside your limits? Check its steps where the tool shows them.
• Are gaps marked "not found" rather than filled with guesses?
• Is anything now saved in memory or the project that should not be there?

Common mistakes

• Handing over the task in one line. A vague task leaves the agent to guess. Write the brief.
• Skipping the plan. Reading a plan takes a minute and catches a wrong direction early.
• Leaving out the limits. Some tools ask before sensitive actions, but do not rely on that. Write what it must not do.
• Never looking at memory. A wrong detail keeps shaping answers until you remove it. Turning memory off may not delete what is already saved.
• Reviewing only the summary. It can sound right while a number in it is wrong. Open the sources for what matters.