YourAIDirector
Start here

Home / Become an AI Director

Hand a task to an AI agent

Director

To hand a task to an AI agent, write the goal, connect only the access that task needs, and keep sending or paying behind your approval. An agent can take the next step on its own. A chat waits for your next message. A custom GPT, a project, and a skill are saved instructions, not a worker. Check the work in phases before you use it.

Next step: copy the first prompt that matches your task. Fill in the brackets. Send only that.

A prompt is the message you paste. An approval gate is a stop you control. A checkpoint is a pause after one phase. School pages teach. They do not do the homework. Follow your teacher's rules.

Last updated

How do you do it?

Do the steps in order. Each line is one action, and you stop at the end of that action.

  1. 1 Write the goal and the stop. One sentence for what done looks like. Then name what the agent must not send, pay, delete, or invent. A chat answers and waits. An agent can take the next step without a new message from you.
  2. 2 Connect only the access this task needs. One folder, one app, or the page you are already on. Leave the rest off. A connector or a connected app is a door into email, files, or a calendar. Open the door this task needs, and no other door.
  3. 3 Keep sending and paying behind a yes you can read. If the product pauses, read the action and use the on-screen approval. If it does not pause, your message is the stop: do the draft, then stop before you send, pay, book, or delete. A scheduled task can run while you are away, so put that stop in the task instructions.
  4. 4 Check one phase, then the next. Read what it did. Keep it, change it, or stop. Do not let phase 2 start on a phase 1 you have not read. The same pause is on Check each phase before the next one.

Work

What prompt fits “A weekly digest that does not send itself”?

You want a short brief from mail you are allowed to use. You still press send.

Do not send, forward, delete, or pay for anything.

Goal: a 10-line digest of mail from [one sender or one subject] since [date].
Access: only that mailbox, if a connector is already on. Do not connect a new app. Do not open other folders.
Done looks like: bullets I can read, each one tied to a subject line I can find.
If you cannot see the mail, say so and stop. Do not invent messages.

When the digest is done, stop. I will decide what to reply to. Do not draft a send until I approve this digest.

School

What prompt fits “A study block, not a finished assignment”?

The agent can quiz you. It cannot write the work you turn in, and it cannot email a teacher.

This is for learning. Do not write anything I will turn in. Follow my teacher's rules.

Do not email anyone. Do not open a shared drive. Do not write a paragraph I could turn in.

Goal: a 25-minute study block on [topic] from these notes only:
[paste notes you are allowed to use]

Phase 1: five questions. Then stop.
I will answer in my own words.
Phase 2 waits until I paste my answers. Then mark what I missed. Do not rewrite my answers into an essay.

Life

What prompt fits “A Saturday plan that does not book or pay”?

The agent can draft an order for the day. You check prices and you pay.

Do not book, pay, message a host, or create a calendar event.

Goal: a Saturday plan in [city] for [who].
Facts: we leave at [time]. We must [must-do]. Budget I will check myself: [amount].
Access: the page I have open, if I pasted it. Do not connect email or a bank.

Give me the order of the day and the list of pages I should open to check hours and prices. Mark every price as CHECK THIS. Then stop.

What mistakes should you avoid?

These are the misses that waste the result. Skip them before you send anything.

  • Saying “handle my email” with the whole inbox connected.
  • Typing a password, a card number, or an ID number into the task so the agent can log in.
  • Leaving a scheduled task allowed to send mail while you are away.
  • Treating a custom GPT, a project, or a skill as if it had checked the facts.

More in this section: Approve the plan before it continues · Check each phase before the next one · Set custom instructions so answers stop being generic

Practice on a real task: Use AI at work · Use AI for school the right way · Keep your data private with AI · Make a project plan with AI

Also on this site: ChatGPT · Claude · Gemini · Copilot · Start here

Last updated · All director guides

Where does this come from?

These are public pages. The note after each link is what that page says, in plain words.

  • OpenAI Help, “ChatGPT Work and Codex”. OpenAI says Chat is for fast questions, and Work is an agent for longer, multi-step work and finished deliverables. Work is on the web and in the mobile app for eligible paid plans. You can review progress, answer questions, change direction, and approve important actions. On the web and on mobile, if an action needs approval, you use the on-screen control. Spoken approval is not supported. If your account does not show Work, use the same stops in a normal chat.
  • OpenAI Help, “ChatGPT agent”. OpenAI’s help says ChatGPT agent is no longer available, and to use ChatGPT Work for longer, multi-step tasks. The rest of that page still describes the old agent. Follow the notice at the top.
  • OpenAI Help, “Scheduled tasks in ChatGPT”. ChatGPT can run a task once, on a schedule, or when a supported app event happens. An action that sends a message or changes data outside the chat may require approval, and the task pauses until you review it. OpenAI says not to put health details, financial details, passwords, or account numbers in the task title or instructions. A task you create inside a project cannot access files uploaded to that project.
  • Anthropic Help, “Get started with Claude Cowork”. Cowork can take a multi-step task and keep working. It is on paid Claude plans. On some accounts the message box no longer shows a separate Chat and Cowork switch, because those are becoming one Claude. Manual mode asks before an action. For money, a message sent as you, or an important file, Anthropic says to stay close or switch back to Manually approve. Claude asks before it permanently deletes a file. You choose which connectors it can use.
  • Anthropic Help, “Use connectors to extend Claude’s capabilities”. A connector lets Claude read and act in another service, using the permissions of the account you connect. Anthropic says to connect only services you trust and need. In a chat you can turn a specific connector on. On Team and Enterprise plans, an owner can set a tool to Always allow, Needs approval, or Blocked.
  • Google Chrome Help, “Customize your Gemini in Chrome experience”. Gemini in Chrome can read the page you are on. A skill is a saved prompt you run by typing / in Ask Gemini. Connected apps, such as Gmail or Calendar, are switched under Personal Intelligence, and that switch applies to Gemini Apps, not only to Chrome. Google says Gemini in Chrome is not available to everyone yet.
  • Microsoft Support, “Transparency Note for Microsoft Copilot (for individuals)”. Microsoft says Copilot Cowork carries out a task you describe, including documents, email, meetings, and files, and that you review and approve sensitive actions before they happen. The same note says to be careful with an agent that could share personal data, and to watch the result. Features vary by subscription. If you do not see Cowork, use the same stops in a normal chat.

What else do people ask?

Open a question. The short answer under it is the one to use.

What is an AI agent, compared with a chat?

A chat answers and waits for your next message. An agent can plan a next step and take it, in a product such as ChatGPT Work, Claude Cowork, or Copilot Cowork, when your account shows that control. You still set the goal and you still check the result.

What is a custom GPT, a project, or a skill?

Those are saved instructions, sometimes with files or a connected app. A custom GPT is a ChatGPT setup for one purpose. OpenAI says new GPT creation is not available on personal accounts, and that GPTs are planned to move to plugins. A project keeps chats, files, and instructions together. A skill is a saved way of working: in Claude, a skill adds a behavior; in Gemini in Chrome, a skill is a prompt you run on the page you are viewing. None of them is a person, and none of them replaces your check.

What if I do not see Work, Cowork, or Gemini in Chrome?

Use a normal chat and the same four stops. The product name is optional. The stop before send, pay, or delete is not.

Will the product always ask before it sends or spends?

No. OpenAI says an action that sends or changes outside data may require approval. Anthropic’s Manual mode asks, and Auto mode does not stop for every step. Microsoft says you review and approve sensitive actions in Copilot Cowork. Write the stop into the task yourself, and read any on-screen approval before you allow it.

Can an agent do my homework?

No. It can quiz you and organize your notes. You write anything you turn in. Follow your teacher’s rules.