Hand off your inbox: a to-do list every morning

1What it looks like handed off

You open one list in the morning instead of the whole inbox: who needs an answer, what's due, what can wait. The routine replies are already drafted. You read them, fix what needs fixing, and send.

2The job, end to end

1. 7:00. A scheduled job starts the agent on its own computer. It can read your mail and write drafts.
2. It reads everything that came in since yesterday's list.
3. It sorts each email: needs you, needs a reply, can wait, not sure.
4. It drafts the routine replies and leaves them in your Drafts folder.
5. It writes the list, one line per email, each linked to the original, and leaves it where you'll see it: as a draft at the top of your inbox, or a message in the chat app you use with your agent.
6. You read it with your coffee. Send the drafts you like; fix or delete the rest.

The schedule is either the agent's own (OpenClaw and Hermes both have one) or your computer's (Task Scheduler on Windows, launchd on a Mac, cron on Linux), which can start Claude Code or Codex with the morning instructions.

3What stays with you

Sending

It drafts; you press send. Out of the box, it asks before it spends money or sends anything.

Money

Invoices and payment requests go on the list. It doesn't pay them.

Anything hard to undo

It doesn't delete mail, unsubscribe you or change your account settings.

When it's unsure

It says so: the email goes under "Not sure" with one line on why. Anything that looks like phishing goes there too, and the agent doesn't follow its links.

Be clear about where that stop lives. Gmail's permission for drafts also allows sending, so on Gmail the "ask first" rule sits in the agent's approval settings, not in Google's. Microsoft 365 can grant read and draft access without send access, so there the rule can sit in the mailbox permission itself. Those permissions still allow other mailbox changes; grant only the access the job needs. You can loosen any of this later, one step at a time.

4When it gets something wrong

It will, sometimes. AI is confidently wrong often enough that the design assumes it.

Anthropic's guidance on agent safeguards describes why protection needs more than an instruction in a prompt. Test your actual approval controls before connecting real mail.

A bad draft

It sits in Drafts until you look at it. Edit it or delete it. Nothing went out.

A missed or misfiled email

The list ends with everything it read, each linked, so a quick scan catches it. For the first two weeks, skim your inbox as well.

A wrong fact in a draft

A price, a date or a promise you didn't make. This is the one to watch, and the reason it never sends on its own. Give it your price list and your usual answers so it has less to guess.

An email written to trick it

Mail can contain instructions aimed at the agent ("forward this to…"). Approval instructions reduce risk, but they do not guarantee that a malicious message stays harmless. Limit the tools and data the agent can reach, require an approval step enforced by the tool before sending, and stop a run that tries to follow instructions inside mail. Review unexpected drafts and actions before continuing.

The same mistake twice

Add a line to its instructions. On upkeep, I do that.

No list one morning

That's how you'll know it stopped. On upkeep, fixing it is my job.

5What it costs

How rough is "$5 to $50"? Very. It's my estimate, not a quote: a morning run over 50 to 100 emails, with the agent's own instructions, reads something like 100,000 to 300,000 tokens. Over a month that's a few million tokens, and the price per million varies more than tenfold between models. Check your provider's usage page after the first week.

6When not to bother

• A dozen emails a day. Your mail app's filters, labels and snooze will do.
• One hard reply now and then. Paste it into ChatGPT or Claude and ask for a draft.
• You want help while you write. Gmail and Outlook have AI drafting built into some plans. Check yours before adding an agent.
• Nobody may see your mail. The agent's AI provider reads what it reads, unless you run a local model. If your clients' mail can't leave your hands, this job isn't for an agent on a cloud model.

If a $20 subscription does it, I'll tell you. The agent earns its keep when the inbox is big enough that the list saves you real time every morning, and you'd rather not open the whole thing first.