YourAIDirector
Start here

Home / Become an AI Director

Keep the send button

Director

To keep the send button, let the chat draft, research, and prepare, then stop it before anything leaves your hands. Send, pay, sign, delete, cancel, and publish each need their own yes, after you have read that exact action. Looks good is not permission. Approve the plan is the stop before more writing. Hand a task to an agent is the stop when a product can keep going. This page is the everyday habit for the last click.

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 Say the hook out loud. Looks good is not permission to act. The chat may draft, research, and prepare. Send, pay, sign, delete, cancel, and publish stay behind a yes you give to that one action, after you have read it. A plan you already approved is not a yes to send.
  2. 2 Make it list the irreversible steps. Before any draft is finished, tell it to list every action that would send, pay, sign, delete, cancel, accept an invite, or publish. Each line needs the exact thing it would do. If there is no such action, it writes NONE. It does not bury the action inside a cheerful summary.
  3. 3 Approve one line at a time. Reply with the line number and the word yes, no, or change. Yes on line 1 does not cover line 2. If a product shows an approval card, read the card. OpenAI says a card may appear before a supported action, and that an action which sends or spends may need review. Anthropic says Gmail send asks by default, and that Duet, if you have that preview, waits before send, pay, sign, or delete. Google’s Help me write page describes a draft in the compose window. You still press send in Gmail. If you cannot find a wait setting, do not invent one. Do the last click yourself.
  4. 4 Leave the button with you. Copy the draft into the mailbox, the form, or the calendar, or stop at the card and read it. Then you send, pay, sign, delete, cancel, or publish. A scheduled task can run while you are away, so put the stop in the instructions. OpenAI says a send may pause for approval. It does not say every send will.

Work

What prompt fits “A reply that is ready, and still not sent”?

The weak version is "Handle this email." That can mean draft, send, and file. The directed version asks for a list of actions and waits. You still press send. This is the everyday habit. Approve the plan is the earlier stop, when the job is still a plan.

Do not send, forward, delete, or schedule anything.

Goal: a reply I might send later, about [subject].
Facts I typed: [three facts].
Reader: [name and role].

First, list every irreversible action this task could include. Use these labels only if they apply: SEND, PAY, SIGN, DELETE, CANCEL, ACCEPT INVITE, PUBLISH. If none apply, write NONE.
Wait. I will reply yes, no, or change to each line. A later "looks good" is not a yes.

You may draft only after I approve DRAFT and I do not approve SEND.
The draft is under 120 words. Then stop. I will paste it into the mailbox myself.

School

What prompt fits “A study note that cannot email the teacher”?

The weak version is "Email my teacher that I understand the chapter." That is a send, and it can become the assignment. The directed version lists the action and stops. You write anything you turn in.

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

Do not email my teacher. Do not submit anything. Do not delete a file.

Goal: help me see if I understand [topic].
My notes: [a few lines you wrote].

List every action that would send, submit, or publish. Label each SEND or PUBLISH. If the honest list is NONE, write NONE.
Wait for my yes, no, or change on each line. Do not treat "looks good" as permission.

Then ask me one question. Wait for my answer in my own words. Do not write the email. Do not write a paragraph I could turn in.
Then stop.

Life

What prompt fits “A bill, a calendar invite, and a post”?

The weak version is "Take care of this." That can mean pay the bill, accept the dinner, and post the photo. The directed version splits them. You tap each one yourself.

Do not pay, book, accept or decline an invite, delete mail, or publish a post.

What I might do later, typed by me:
- Bill: [who, and the amount I copied, or AMOUNT NOT IN MY NOTES]
- Invite: [title and time I can see, or NOT IN WHAT I PASTED]
- Post: [the sentence I wrote, or NONE]

List these as separate lines: PAY, ACCEPT INVITE, PUBLISH. Put NONE beside any line I did not actually describe.
Wait. I will answer yes, no, or change to each line by itself.
Do not take an action I have not named. Do not invent a price, a time, or a caption.
A single "looks good" does not approve the list.
Then stop.

What mistakes should you avoid?

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

  • Saying looks good, or sounds right, and treating that as a send.
  • Approving the plan, then assuming the agent may also pay or email.
  • Turning on Allow all actions because the card was annoying. OpenAI says that option carries elevated risk.
  • Leaving a scheduled task allowed to send while you are away, with the stop left out of the instructions.
  • Assuming every account has Claude Duet. It is a US Pro, Max, and Ultra research preview, by waitlist, and not on Team or Enterprise yet. If you do not have it, you still keep the button.
  • Assuming Gemini sent the mail because it wrote a draft. Google’s Help me write page describes a draft in the compose window.
  • Using a school prompt to get the email or the assignment sent. You write what you turn in.

More in this section: Approve the plan before it continues · Hand a task to an AI agent · Ask for a plan before the work · Ask for a morning brief

Practice on a real task: Catch up on your email and calendar with AI · Keep your data private with AI · Reply to work emails · Write a difficult work email · Plan your day with AI

Also on this site: Work · Life · School

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, “Managing app permissions in ChatGPT”. OpenAI says app permissions control when ChatGPT can read a connected account or take an action. Always ask means it asks before reading or making changes. Allow read actions can read without asking and asks before changes. Allow all actions can perform supported actions without additional approval prompts, and OpenAI says that option carries elevated risk. Before a supported action runs, ChatGPT may show an approval card. Sending an email or making a purchase is an example of an action that may require additional review. The standard account-wide selectors do not offer Allow all actions. It may be available for an individual app or connected account. If your account does not show these menus, your message is still the stop.
  • OpenAI Help, “Scheduled tasks in ChatGPT”. OpenAI says an action that sends a message or changes external data may require approval, and that if approval is required the task pauses until you review it. Do not put health information, financial details, passwords, or account numbers in the task title or the instructions. May require is not a promise that every send will wait.
  • Anthropic Help, “Use Google Workspace connectors”. Anthropic says Claude can send, reply to, and forward Gmail, and that by default it asks for your approval before each of those actions. On Team and Enterprise plans, owners decide whether members can allow those actions to run without asking each time. The same page says that by default each action Claude takes on your behalf requires your explicit approval. Calendar tools on that page can create, update, and delete events and respond to invitations. Read the approval before you accept one.
  • Anthropic, “Introducing Claude Duet” (October 23, 2026). Anthropic’s research preview of Claude Duet is a waitlist, starting with US Pro, Max, and Ultra subscribers, on iOS, Android, and claude.ai. The article says Duet reads, drafts, and prepares on its own, and that before it sends a message, makes a payment, signs a document, or deletes anything, it shows what it is about to do and waits for your approval. Trading securities and moving money between accounts are off-limits even with approval. If you are not in the preview, you do not have this app. The habit on this page does not require it.
  • Anthropic Help, “Get started with Claude Duet”. The help page says Duet is designed to show you what it is about to do and await your approval before sending a message, making a payment, signing a document, or deleting anything. If you cannot approve right away, it can finish the rest of the task and bring the approval back later. It is for Pro, Max, and Ultra plans in the US, by waitlist, and not on Team or Enterprise yet. It can still misread a thread. Review the approval before you confirm it.
  • Google Gmail Help, “Draft emails with Gemini in Gmail”. Google describes Help me write as a way to generate or refine a draft in Gmail’s compose window. The steps end at creating and refining the draft. The page does not say Gemini sends the message. Sending stays on Gmail’s send button. The same page says not to rely on Gemini features as medical, legal, financial, or other professional advice, and that they may suggest inaccurate information. It also says Help me write can use details from your other emails and Drive files to personalize a draft. Read the draft, and use Sources if you need to see what it pulled in.

What else do people ask?

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

What does keep the send button mean?

The chat may draft, research, and prepare. You send, pay, sign, delete, cancel, and publish. Each of those needs a yes to that action after you have read it. Looks good is not that yes.

How is this different from approving the plan?

Approve the plan is the stop before the chat continues writing. This page is the stop before an action leaves your account. You can approve a plan and still refuse the send.

How is this different from handing a task to an agent?

That guide is for a product that can take the next step on its own: the goal, the access, and the phase check. This page is the same last click in an ordinary chat, a connected mailbox, a calendar, or an agent. Use both when the product can keep going.

Will the product always wait?

No. OpenAI says a send or a change outside the chat may require approval, and that Allow all actions can skip extra prompts. Anthropic says Gmail send asks by default, and that owners on Team and Enterprise can allow actions without asking each time. Claude Duet, if your account has that research preview, is described by Anthropic as waiting before send, pay, sign, or delete. Google’s Help me write page describes a draft, not a send. If you do not see a wait, you do the click.

What is Claude Duet?

Anthropic’s October 23, 2026 article describes a research preview that starts with US Pro, Max, and Ultra subscribers, by waitlist. It says Duet shows the action and waits before it sends, pays, signs, or deletes. The help page says it is not on Team or Enterprise yet, and that it can still misread a thread. If you are not invited, you do not have it. The morning-brief guide is the standing note. This page is the button.

Will this do my homework or send the email?

No. It can ask you a question and wait. You write anything you turn in, and you press send yourself. Follow your teacher’s rules.