AI Prompts for a Shared Inbox: 15 That Run a Team's Email (2026)

Most AI email prompts are written for one person's inbox. A shared inbox has different problems. The expensive ones are not "write this reply better": they are three people drafting the same answer, a thread nobody owns, a handoff that loses its context, and five teammates answering in five voices. The prompts below are built for those failures. The first twelve work by copy-paste in Claude, ChatGPT, or Gemini. The last three do something copy-paste cannot: they run a connected inbox directly, through MCP, and report back.

One rule before the prompts, because it is the difference between a useful system and a mess: everything here drafts, sorts, and summarizes. A human reviews and sends. If you want the setup that lets the last three prompts run on their own, our guide to connecting a shared inbox to Claude covers it in about ten minutes.

Why shared inbox prompts are different

A personal-email prompt has all its context in the thread you paste. A team prompt has context the thread does not carry: who owns what, what was promised by a teammate last week, which answers are policy and which are judgment. Good team prompts do two things the personal kind never need: they ask the AI to make ownership visible, and they feed it the team's voice instead of yours. Every prompt below does at least one of the two.

Triage prompts: clear the queue, in order (1-4)

1. The morning sweep. This one stops you reading forty threads in arrival order when three of them actually matter.

Works in Claude, ChatGPT, or Gemini

Here are the subject lines and first lines of everything unanswered in our support inbox this morning: [paste]. Sort them into three groups: needs a reply within the hour, needs a reply today, can wait or close. One line of reasoning per thread, no more.

Copy

2. The unanswered-and-aging list. This catches the thread that quietly turns three days old.

Works in Claude, ChatGPT, or Gemini

From this list of open threads with their last-reply dates [paste], list every thread where the last message is from the customer and is older than 24 hours. Rank by how annoyed the customer will be, judging from their last message.

Copy

3. The duplicate-cluster catch. This stops three teammates writing three different answers to the same incident.

Works in Claude, ChatGPT, or Gemini

Here are today's incoming subjects and senders [paste]. Group any that look like the same underlying issue reported by different people, and draft one canonical answer we can adapt for each.

Copy

4. The VIP filter. This keeps your biggest account from waiting behind newsletter unsubscribes.

Works in Claude, ChatGPT, or Gemini

Here is our list of priority customers [paste] and here is the unanswered queue [paste]. List every thread from a priority customer, oldest first, and flag any where the tone suggests frustration.

Copy

Ownership and handoff prompts: who has this, and what do they need to know (5-8)

5. The assignment suggestion. Without this, everything defaults to whoever opens the inbox first.

Works in Claude, ChatGPT, or Gemini

Based on these open threads [paste] and what each teammate handles (Ana: billing. Sam: technical. Jo: everything else), propose an owner for each thread with one line of reasoning. Format as a list I can act on in one pass.

Copy

6. The handoff summary. This replaces "can you take this one?" followed by the new owner reading a 30-message thread from scratch.

Works in Claude, ChatGPT, or Gemini

Summarize this thread [paste] for a teammate taking it over: what the customer wants, what we have already promised, what is still open, and the one thing they must not contradict. Five lines maximum.

Copy

7. The escalation note. This stops escalations arriving with no context, where the manager's first move is asking the customer to repeat themselves.

Works in Claude, ChatGPT, or Gemini

Turn this thread [paste] into an escalation note for my manager: the issue in one sentence, what we tried, why I am escalating, and the decision I need from them. Keep it under 100 words.

Copy

8. The out-of-office reroute. This is for the week a teammate is away and their open threads would otherwise sit unclaimed.

Works in Claude, ChatGPT, or Gemini

Sam is out until Monday. Here are Sam's open threads [paste]. For each: can it wait for Monday, or does it need a new owner now? If it needs one, say who between Ana and Jo, based on what the thread is about.

Copy

Team-voice drafting prompts: one voice, many hands (9-12)

9. The house-voice reply. This is the fix for five teammates writing in five styles to one confused customer.

Works in Claude, ChatGPT, or Gemini

Here are three replies we consider our best tone [paste]. Learn the voice: how we open, how we apologize, how we say no. Now draft a reply to this thread [paste] in that voice. Do not exceed 120 words.

Copy

10. The de-escalation draft. This stops the two classic mistakes with angry email: answering in kind, and answering in corporate.

Works in Claude, ChatGPT, or Gemini

This customer is angry [paste]. Draft a reply that names their real complaint in the first sentence, gives the concrete fix and its timing, and offers one thing we can do today. No apology stacking: apologize once, properly.

Copy

11. The policy-consistent no. This prevents one teammate granting the refund another refused last week.

Works in Claude, ChatGPT, or Gemini

Our refund policy: [paste policy]. This customer is asking for an exception [paste thread]. Draft a reply that holds the policy, explains the reason in plain words, and offers whatever the policy DOES allow. Flag for me if you think this case is genuinely borderline.

Copy

12. The promise ledger. This finds the commitments made in threads that nobody wrote down.

Works in Claude, ChatGPT, or Gemini

Read these five threads [paste]. List every commitment we made: what was promised, to whom, by when, and whether the thread shows it was done. Table format.

Copy

The prompts that work the inbox themselves (13-15)

Everything above needs you to paste. These three do not. Connect your shared inbox once through MCP, and the assistant reads the queue, acts on it, and shows its work with real data. They are written against Drag's MCP server and run verbatim; the tool names in the results (search_threads, get_response_times, move_thread) are the server's actual tools doing the work.

13. The self-serve morning sweep. Prompt 1, without the pasting.

Claude

Look at our support board. Find every thread where the customer is waiting on us, group by urgency, and tell me the three I should answer first and why. Then draft a reply for the most urgent one and leave it for my review.

Copy

What happens: the assistant calls search_threads on the connected board, reads the waiting threads, ranks them, and drafts into the thread it chose. The draft waits for a human. Nothing sends itself.

14. The Friday report that writes itself.

Claude

Pull our response times for this week and compare them to last week. Which day was worst, which teammate is carrying the most volume, and did anything sit unanswered for more than a day? Write it up as our Friday review, in plain sentences.

Copy

What happens: get_response_times and get_daily_activity return the real numbers; the write-up cites them. The report is grounded in the inbox, not in what anyone remembers.

15. The queue gardener.

Claude

Find every thread on the support board that has been resolved but is still sitting in the In Progress column, and move them to Done. List what you moved so I can spot-check.

Copy

What happens: search_threads finds the stale cards, move_thread files them, and the list comes back for review. Five minutes of weekly housekeeping becomes one sentence.

What the difference looks like in practice

Prompt 6 with a real thread pasted returns a decent summary. The same request against a connected inbox returns the summary plus what the thread's labels say, who touched it last, and how long it has been open, because the assistant can see the board, not just the text. That gap exists because pasted text never carries the context around it. Closing it is why we built our own MCP server rather than waiting for someone else to.

What the last three prompts need to run, and why a normal Gmail connection is not enough

The copy-paste prompts work with any inbox and any assistant today. The last three need two things: a shared inbox that is more than a mailbox, and an MCP server that lets your assistant operate it.

Here is why the ordinary route falls short. If you connect your personal Gmail to Claude or ChatGPT, the assistant sees one person's mail and acts as that one person. It cannot see who owns a thread, because plain Gmail has no concept of ownership. It cannot tell a thread in progress from one that is finished, because a mailbox has no statuses. And whatever it does is invisible to your teammates, which on a shared address is exactly the problem you were trying to solve.

A shared inbox adds the missing structure, and that structure is what the prompts actually use. In Drag, our shared inbox for Gmail, every thread has an owner, so "who is carrying the most volume" is a real question with a real answer. Threads live in columns like New, In Progress, and Done, so "find resolved threads still sitting in In Progress" means something. Labels carry the triage taxonomy. Response times are measured continuously, so the Friday report cites numbers rather than impressions.

The MCP server is the second half: it exposes all of that to your assistant as 47 tools across 12 categories, searching threads, drafting replies, moving cards, reading the analytics. Two properties matter more than the tool count. Every action is attributed and visible to the whole team, so an assistant working the queue is governed the same way a teammate is. And drafting stays separate from sending, so the human-reviews rule from the top of this page is enforced by the system, not by discipline.

Drag costs $12 to $24 a seat, with AI included from $18 and nothing metered, after a free 7-day trial. The setup guide takes about ten minutes, and the assistant you already pay for, Claude, ChatGPT, or Cursor, becomes the teammate who never lets the queue age.

Frequently asked questions

Do these prompts work in ChatGPT, Claude, and Gemini?

Prompts 1 to 12 do: they are plain text plus whatever you paste, and all three assistants handle them well. The last three need an assistant that supports MCP connections, which today means Claude, ChatGPT via Developer Mode, and Cursor, connected to an inbox that exposes an MCP server.

Can the AI answer customers directly with these prompts?

No, and that is deliberate. Every prompt here ends in a draft or a report for human review. Autonomous customer-facing agents are a different product category with different failure modes; if that is what you are shopping for, our AI support agent resolution rates audit is the honest place to start.

What does it mean for a prompt to run on a connected inbox?

A prompt an assistant can act on through a live connection instead of pasted text. MCP is the open standard that lets AI tools call other software's functions. Connected to a shared inbox, "find every waiting thread" stops being a request for a summary of what you pasted and becomes an instruction the assistant executes against the real queue.

Do I need Drag to use this page?

For prompts 1 to 12, no: they work with any inbox and any assistant. Prompts 13 to 15 need an inbox with an MCP server, and Drag is ours: $12 to $24 a seat, AI included from $18, free 7-day trial.

How is this different from your ChatGPT and Claude email prompt guides?

Those cover one person's email: replies, follow-ups, tone. This page covers the problems that only exist when a team shares one inbox: who owns a thread, how handoffs keep their context, how five people write in one voice, and how the reporting writes itself. The prompts are different because the problems are.

Related guides: personal-email prompts and the MCP setup

For personal-email prompts, 50+ ChatGPT prompts for email and the 15 Claude email prompts are the deeper libraries. For the connection itself, connect a shared inbox to Claude is the setup walkthrough, and connect Gmail to Claude, to ChatGPT, or to Gemini cover the personal-account routes. For what else MCP can reach, the best Claude MCP servers for business audits the field.