Build an Email Bot That Drafts, Never Sends

Email is the first Grok Bot worth building because the job is obvious, the volume is daily, and the failure mode is obvious too. A useful email bot drafts and sorts. A dangerous one sends.

This page assumes you already have one bot standing from part 1. If you do not, go do that first. Do not bolt Gmail onto a bot that is also writing code or paying invoices. Inbox gets its own teammate.

Why a dedicated Inbox bot

Mixing jobs is how you get a draft to a customer that quotes a number from a different thread, or a "looks good" that was meant for a school form. One bot, one mailbox, one voice. You can add a Follow-ups bot later and put them in the same thread. That is a later part of this series. Not today.

Name it Inbox. Put the mailbox in the name if you have more than one: Inbox-Work, Inbox-School. The name is a constraint. The bot should be able to say "I am Inbox-Work, I do not touch the family Gmail."

The job brief you actually paste

Send this, then edit the bracketed lines so they are true for you. A brief that still says "your company" on day three is why drafts sound like a vendor demo.

You are Inbox. You work in Gmail only. You never send.

Goal: I spend 15 minutes on mail each morning instead of 90. You do the reading. I do the decisions.

Every working morning, before I ask:
1. Unread count, plus how many can wait
2. The five threads that need me today, one sentence each, oldest first inside that five
3. An unsent draft sitting on each of those five
4. A separate "do not miss" list: bills, legal, school, medical, anything with a date

Rules:
- Never send, never delete, never archive a thread I have not seen
- Never invent a fact, a price, a date, or a name you cannot see in the thread
- If a thread is angry, legal, or about money, draft nothing. Summarise and wait
- Match my voice from the last three emails I actually sent in that thread
- If you cannot open Gmail, say so. Do not work from memory of yesterday

My voice: short, direct, no exclamation points, sign-off is just my first name.
People who always get a human reply: [names].
People I am fine delaying a day: [names or labels].
Never-send list: anything to [lawyer / accountant / school principal / that one vendor].

The never-send list is the most important paragraph. Write it while you are calm. You will not write it well at 11pm after a bad email.

Sign it into Gmail. You stay for 2FA

Ask the bot to open Gmail in its browser and stop at the login screen. You complete the Google sign-in and the second factor. Do not paste a password into the chat. Do not disable 2FA to "make it easier for the bot."

Use the mailbox you mean. A bot on the wrong Google account will cheerfully triage someone else's life and you will not notice until a draft mentions a thread you have never seen.

If you use Google Workspace, prefer the work account and keep personal mail off this bot. If you only have one Gmail, say that in the brief so it does not try to "find your work inbox."

xAI describes this as logging the bot in once so it can use the site like you would, including tools with no API. Gmail is exactly that case. The session will persist on the bot's computer. Treat that like leaving a logged-in laptop in a locked room: fine for a contractor, not fine if you also keep password-reset mail in the same inbox without a rule.

Week one: triage only

Do not let it draft on day one. A week of triage teaches you whether it can see the mailbox, whether it invents urgency, and whether its "five that need you" matches yours.

A good morning brief looks like this:

• 47 unread. 31 are newsletters or receipts. 16 need a look.
• Need you: (1) insurer asking for a photo by Friday (2) school trip form (3) vendor waiting on a yes (4) refund that bounced (5) intro from someone you asked to meet.
• Do not miss: insurer Friday deadline.
• I have not drafted. Say the word and I will.

A bad brief looks like this:

• You have some emails.
• A lot of them seem important.
• I went ahead and replied to three so you would not have to.

If you get the bad one, the brief failed or the bot ignored it. Do not "just this once" keep the sends. Make it show you the three drafts and you decide whether any of them actually went out. Then tighten the rule: "If you sent anything, say so in the first line, every time, even if I told you not to."

Week two: drafts that stay unsent

Once the triage matches your judgment three mornings in a row, let it draft. The draft lives in Gmail as a draft. Not a sent item. Not a scheduled send. You open it, edit two words, and you hit send.

Tell it how to draft, not just that it should:

• One screen or less. If the thread needs a page, it needs you, not a draft.
• Answer the question that was asked. Do not add a paragraph of "happy to help."
• Pull names, dates, and numbers only from that thread. If a price is not in the thread, leave a [PRICE?] token and stop.
• If the other person is upset, you write the first sentence and stop. You do not try to win.

Read every draft for a week. You are not only checking tone. You are checking whether it replied to the latest message or to a quote from 2024. Bots that can see a whole thread will still answer the wrong one if you do not say "reply to the newest message only."

Buckets that actually cut the pile

Do not invent a 20-label taxonomy. Four buckets cover most inboxes:

• Needs me today. A decision, a form, a date.
• Waiting on them. You already replied. The bot tracks, it does not poke yet.
• Can wait. Real mail, not today.
• Noise. Receipts, newsletters, notifications. Summarise weekly, not daily.

If your host lets the bot apply labels, let it label Noise and Waiting. Do not let it label Needs me today without putting those threads in the morning brief. A hidden "urgent" label you never see is how deadlines die.

Memory worth saving

After a few corrections, tell the bot what to remember. Keep this short and true:

• Who always gets a human reply, by name.
• The sign-off you actually use.
• Words you never want in a draft (circle back, looping in, great question).
• The never-send list.
• Recurring mail that looks urgent and is not (the weekly vendor newsletter with "ACTION REQUIRED" in the subject).

Do not dump your life story into memory. A bot that "knows" too much will start using a fact from a school thread in a work draft. If a fact is only true in one mailbox, it belongs in that bot's brief, not in a shared brain.

The morning pass, once triage is boring

When the brief is stable, you can turn it into a routine: you do the morning pass once with the bot watching, it saves the steps, it runs them the next day and leaves the brief for you. That is the next lesson. You do not need it to get value this week. A message that says "run the morning brief" is enough until the steps stop changing.

Approvals stay with you. iOS is useful here. Read the five threads and the drafts on your phone. Send from the phone only after you have opened the draft. Never approve a send from a lock-screen summary.

What this bot must never do

Write these as standing rules, not vibes:

• Never send. Not even "quick yes." Not even to yourself as a test, unless you said so in that message.
• Never reply-all unless the draft is a reply-all and you approved that.
• Never forward a thread outside the mailbox.
• Never download a mystery attachment and run it.
• Never use a password from mail to log into something else.
• Never mark a legal, school, medical, or money thread as Noise.

Part 6 is the longer version of this list. If you only remember one line, remember this one: the bot's job is to make the send button easier for you to press, not to press it.

Done when

The email bot is real when all of this has happened on a normal weekday:

• It opened the right Gmail.
• The morning brief named the same five threads you would have picked.
• Five drafts sat in Gmail, unsent, and you used at least two of them after a light edit.
• Nothing left the mailbox without you.

When that is boring, the next lesson is teaching the morning pass as a routine. That part is next in the series.