Last updated:
Quick answer
Tell the agent, in its standing instructions, to finish every outgoing action as a draft and show it to you with a one-line summary, then wait. Keep the tool's own approval setting on as a second lock, and test it with a fake send before trusting it.
Free PDF
Get this setup as a PDF
Confirm your email and we'll open a print-ready copy of this setup you can save as a PDF. You also get the free starter pack and one setup a week. Unsubscribe anytime.
Already a subscriber? Unlock it on this device
At a glance
All outcome figures are estimates from cited sources, not independent verification by us.
- What it is allowed to do
- Drafts for you
Writes drafts (emails, posts, plans, replies). You review and send. Nothing goes out without you.
- Time saved (estimate)
- About 10 minutes to set up and test.
- Difficulty
- Beginner
- Tools needed
- ChatGPT, Claude
- Typical cost
- Works on free tiers for drafting. Connected-app agents may need a paid plan; prices: Not stated in sources.
Reviewed Nov 2, 2026
Basis (estimate): Estimate only: based on the steps above, not timed.
What does this setup do?
Make your AI agent stop and show you every email, booking or post as a draft with a one-line summary, and wait for your OK before it goes out.
Who is it for?
Anyone whose AI assistant can send, book or post through connected apps, and who wants a human check on everything that leaves the business.
How do you set it up?
Tools
- ChatGPT or Claude with the apps you use connected
- The agent's project or custom instructions
Steps
- Find where your agent keeps standing instructions: project instructions, custom instructions or the agent's settings.
- Paste the approval rule from the prompt below. It turns every outgoing action into a draft plus a short summary of who, what and when.
- Keep the tool's own approval setting on as a second lock. Anthropic says its Google Workspace connectors ask for approval before sending, replying or forwarding by default.
- Test it: ask the agent to email yourself a fake quote. It should stop and show the draft, not send.
- Approve with a clear word, such as SEND, so a vague reply like ok does not count as approval.
- If it ever sends without asking, disconnect that app, tighten the instruction and test again.
Example output
Illustrative. Your results will differ.
- DRAFT 1 of 3: email to dana@example.com
- Subject: Your deck quote
- Goes out: when you reply SEND
- Risk if wrong: wrong price sent to a customer
- Reply SEND to send, or say what to change.
Where to paste this prompt
ChatGPT → Project instructions, or ChatGPT Tasks for scheduled reminders (once a day on Free)
Copy-paste prompt
Adapt bracketed fields to your situation.
Add this rule to your standing instructions and follow it every time:
Before anything leaves my business (email, message, booking, calendar invite, post, payment, refund):
1. Prepare it as a draft only.
2. Show me: who it goes to, the full text, and when it would go out.
3. Add one line: what could go wrong if this is wrong.
4. Wait. Only proceed when I reply with the word SEND for that item.
5. If I reply with anything else, treat it as an edit or a no.
If you are unsure whether something counts as leaving the business, treat it as if it does.Newsletter
Get the free starter pack
Free starter pack with five beginner copy-paste prompts after you confirm, plus one researched setup in your inbox each week. Unsubscribe anytime.
What usually goes wrong?
Where agents tend to stall on this job. More patterns in Where AI gets stuck.
- The instruction lives in one chat, so a new chat forgets it. Put it in project or custom instructions.
- Turning off the tool's own approval setting because the agent asks anyway. Keep both.
- Approving from your phone without reading the recipient line.
Which chat tools work with this prompt?
Paste the prompt into any of these assistants (free or paid tiers).
- ChatGPT
- Claude
Common questions
- How do I stop an AI agent from sending emails on its own?
- Put a stop-and-show rule in its standing instructions and keep the tool's own approval setting on. Then test with a fake send.
- Will it still save me time if I approve everything?
- Yes. Reading and approving a finished draft takes seconds; writing it was the slow part.
- What if the agent ignores the rule?
- Disconnect the app it used, tighten the instruction and test again before reconnecting.
Where should you go next on this site?
What tips make it smoother?
- One approval word per item stops a single yes from releasing a whole batch.
- Ask for the one-line risk note; it makes bad drafts obvious at a glance.
- Re-test after any change to the agent's apps or model.
Where do the claims come from?
Summaries use public sources linked below, not independent verification. How setups are picked.
- Read Anthropic’s write-up (support.claude.com)
Anthropic: Use Google Workspace connectors: Gmail drafting, approval before send, reply or forward by default (read 2026-10-31 through a page fetch tool, not a direct browser open)
- Read OpenAI’s write-up (help.openai.com)
OpenAI: ChatGPT agent: when signed into sites or apps it can access sensitive data like emails, files or account settings and perform actions (read 2026-11-02 from the official help page via search)
- Read OpenAI’s write-up (help.openai.com)
OpenAI: Projects in ChatGPT: keep chats, files and instructions together; owners update instructions and add files (read 2026-11-02 from the official help page via search)
Related setups
- Set guardrails before an AI agent gets access
Write a one-page rulebook for an AI agent before it touches your accounts: what it can read, what it can only draft, what always needs your approval, and what it must never do.
- Morning inbox briefing and multi-inbox triage
Start the day with a short “what needs you” note, and periodically sort multiple inboxes into reply vs FYI vs junk, with drafts for your approval, not auto-send.
- Give each AI agent its own email
Give each AI helper its own sending address, such as ops@ or billing@, so its drafts are easy to tell apart and nothing goes out without you.