The prompt to summarize an email thread asks for four lists, not a summary
Use this prompt for the thread you have to act on. Microsoft's Work Trend Index special report of June 2025 counts 117 emails a day for the average worker, "most of them skimmed in under 60 seconds", and a 60 message thread that moved a decision is the one a skim gets wrong. The prompt asks for what was decided, what is open, who owes what and what is yours, with a quote behind every line.
```
Here is an email thread, oldest message first. Give me four lists:
- Decided: what was agreed, by whom, on what date
- Still open: questions nobody has closed
- Owed: who owes what to whom, with the date written in the thread
- For me: anything asked of me (I am [your name])
Quote the sentence behind every item. Mark a hedged commitment ("should be able to by Friday") as not promised. Skip greetings and thanks.
[paste the thread]
```
The quote line is the safety net. A line with a quote can be found in the thread with Ctrl+F; a line without one is the model inferring, and you cannot check it. The hedged commitment line catches the item a week away makes easy to misread, because "should be able to by Friday" is not a promise. The chart is why the paste needs five minutes of preparation. When each reply quotes the messages before it, a 60 message thread holds 1 + 2 + ... + 60 = 1,830 message copies, 30.5 times the 60 you need. At 120 words a message, about 200 tokens by Anthropic's rule of about 3.5 English characters per token, that is about 370,000 tokens raw against about 12,000 stripped.
Seven more prompts cover one email, an inbox, the reply and Copilot
Seven more prompts cover the other email jobs, each asking for a named output rather than a summary. Paste the email or thread under the prompt. For writing the replies, the email prompt pack has twenty more.
| Job | Prompt to paste |
|---|---|
| One long email | Summarize this email in three lines: what it is about, what it asks of me, and the deadline if there is one. If it asks nothing of me, say so. |
| A thread you half read | I read this thread up to [name]'s message on [date]. List only what changed after it: new decisions, reversals and new asks, with a quote for each. |
| A decisions log | List every decision in this thread in date order. Under each, name any later message that reversed or reopened it, and quote that message. |
| Action items | Make a table of action items with five columns: person, task, owed to, due date as written, quoted sentence. Leave the due date blank where the thread gives none. |
| A pile of separate emails | Here are [N] separate emails. Make one table row per email: sender, subject, what they want from me, deadline, and whether it can wait a week. Sort by deadline. |
| The reply | Using only the items addressed to me above, draft one reply that answers each in thread order, under 150 words, with no apology for the delay. |
| Copilot in Outlook | Summarize my emails from the last week about [project] as four lists: decided, still open, who owes what to whom by when, and anything asked of me. |
The Copilot row extends Microsoft's own example prompt, "Summarize my emails from the last week related to [project or customer name]", from Chat with Copilot in Outlook, which opens from the Copilot button for a work or school account (as of September 2026). The other six run in ChatGPT, Claude, Gemini or any assistant you paste into. Keep the blank due date rule in the action items prompt. A date the model inferred looks exactly like one somebody wrote. If you triage a client's inbox for a living, AI tools for virtual assistants has the sorting prompt that runs before any of these: one line per email, with a label for reply today, reply this week, file or flag.
Why does AI get the order of an email thread wrong?
AI gets the order wrong when the thread is pasted newest first, because a model reads a paste top to bottom and treats the top as the beginning. Outlook sorts that way. Microsoft's conversation view help page says "Within each Conversation, messages are sorted with the newest message on top" and, of classic Outlook, "You cannot change this sort order." Gmail is the opposite: Google's Workspace guide says "the newest message is at the bottom", so a Gmail thread copied from the top comes out oldest first. Any forwarded thread nests newest first, with the earliest message buried at the bottom of the quoted stack.
Paste newest first and the timeline runs backwards. Tuesday's "let's go with vendor B" sits above Monday's "vendor A, agreed", so the reversal reads as the original position and the original decision reads as an objection to it. Dates in the message headers give a model something to catch this with. A thread where people write "as agreed yesterday" under a stripped header gives it nothing.
The fix is mechanical. In Gmail, expand every message in the thread and copy from the top. In Outlook, copy the messages one at a time from the bottom of the conversation up, or paste the lot into a plain text editor and move the blocks. If you cannot reorder, change the first line of the prompt to "newest message first". Telling the model the order is the second best option; giving it the order is the first.
How much of an email thread fits in one prompt?
About 100 to 140 stripped messages fit in one Whizi message as of September 2026, and a raw 60 message thread does not fit at all. One Whizi message is capped at 100,000 characters and one turn at 40,000 input tokens on every model in the catalogue, and the length caps guide quotes each refusal. The character cap lands first: at about 700 characters a message, 100,000 characters is about 140 messages, and sender and date lines pull that toward 100. A raw 60 message thread, with its 1,830 quoted copies, runs past a million characters. The refusal names your count, "over the 100,000 character limit", and it is a refusal, not a trim: nothing is sent.
So paste into a plain text editor first. Delete every line that begins with ">". Delete everything below each "On [date], [name] wrote:" or "From:" header inside a message, which is where the quoted chain starts. Delete signature blocks, legal footers and unsubscribe lines. What is left is the 60 bodies, in order, with a sender and date line above each one, and that is the paste.
Past about 100 messages, split the thread at a natural date. Run the four list prompt on each half, then paste both answers into one turn and ask for a single merged owes list.
How do you check an AI email summary before you act on it?
You check an AI email summary by searching the thread for every quoted sentence and discarding any item whose quote is not there. Each list has one more check of its own:
| List | What each line should carry | The check before you act |
|---|---|---|
| Decided | The decision, who made it, the date, and the quoted sentence | No later message reopened it |
| Still open | The question, who last touched it, and what would close it | Nobody closed it later in the thread |
| Owed | Person, deliverable, to whom, the date, and the quoted sentence | The date is the one in the message, not an inference |
| For me | The ask, who asked, and the quoted sentence | Your name is in that message |
A quote that is not in the thread means the model blended two messages, and a stacked quoted chain gives it the most chances to. Ask again with the message number. A decided item that a later message reopened is the second thing to look for, and it is the one a week away makes you most likely to miss.
Then run the same prompt in a second model and read only where the two lists disagree. Where one lists an open item the other calls closed, open the thread at that message, because one of them misread a reversal. In Whizi, Compare puts two models' answers in side by side columns on Pro and Powerhouse; on Starter, switch models mid-conversation and ask the second model where it disagrees with the first list.
Is there an AI that reads your emails without a paste?
Yes: Gemini in Gmail and Copilot in Outlook both summarize a thread in place, and for a thread that lives in one mailbox they are faster than any prompt. In Gmail, "At the top of the thread, click Summarize this email"; Google says the feature is "available to all Gmail users globally" with the smart features settings on, in some languages only, and that it synthesizes key points and replies and can suggest reminders for tasks with a due date (Gmail Help, as of September 2026). In Outlook for Microsoft 365, new Outlook for Windows, Mac, iOS and Android, "Summary by Copilot" sits at the top of the thread and can include numbered citations that jump to the message they came from (Microsoft Support). Neither needs a paste, so neither has the order trap.
| As of September 2026 | Gemini in Gmail | Copilot in Outlook | A pasted prompt (ChatGPT, Claude, Gemini) |
|---|---|---|---|
| Reads the thread | In place, no paste | In place, no paste | You paste it, oldest first |
| Order and signatures | Gmail handles it | Outlook handles it | Your job, about five minutes |
| Output shape | Key points and replies, reminder suggestions | Summary with numbered citations to messages | The four lists, with quotes |
| A thread that spans Slack, email and doc comments | No | No | Yes, paste all of it |
| A second model's read | No | No | Yes |
| Who can use it | Gmail accounts with smart features on, some languages | Outlook for Microsoft 365, new Outlook for Windows, Mac, iOS, Android | Anyone, free ChatGPT included |
A pasted prompt wins in three cases: the decision moved between Slack, email and a shared doc; you want the four lists with quotes rather than key points; or you want a second model's read. Free ChatGPT is enough for the paste, since OpenAI removed limits on text chats for all users on 6 August 2026 (TechCrunch). For a one mailbox thread you only need the gist of, press the button. If you have to act on it, paste it oldest first and stripped, ask for the four lists, and search for every quote before you reply to anyone.
- Paste the thread oldest first; classic Outlook and every forwarded thread hand it to you newest first
- Strip quoted replies (lines starting with ">") and signature blocks before pasting
- Ask for four lists: decided, still open, who owes what to whom by when, and anything asked of you
- Require a quoted sentence for every item and search the thread for each quote
- Keep one paste under 100,000 characters, roughly 100 to 140 clean messages
- Run the prompt in a second model and open the thread wherever the two lists disagree
- Reply to the items addressed to you first; they are the ones waiting on you
Frequently asked questions
What is a good prompt for summarizing emails?
A good prompt names the output instead of asking for a summary: "Give me four lists: decided, still open, who owes what to whom with dates, and anything asked of me, with the quoted sentence behind every item." Say "oldest message first" above the paste, and add a line telling the model to mark hedged commitments as not promised.
How do I get AI to summarize my emails?
Press the built-in button, or paste. Gmail shows Summarize this email at the top of a thread for accounts with smart features on, and Outlook for Microsoft 365 shows Summary by Copilot. For a thread that spans Slack and email, or one you want a second model to check, paste it oldest first into ChatGPT, Claude or Gemini with the four list prompt.
Can ChatGPT summarize emails for free?
Yes. Free ChatGPT has had no limit on text chats since OpenAI's change of 6 August 2026, so a stripped 60 message thread and the four list prompt cost nothing. Separate limits still apply to files, so paste the text rather than attaching an export.
How long an email thread can AI summarize in one go?
In Whizi, as of September 2026, one message is capped at 100,000 characters and one turn at 40,000 input tokens, which is roughly 100 to 140 clean messages of about 120 words each. A thread pasted raw, with every reply quoting the chain beneath it, is many times larger and gets refused, so strip the quoted replies first, and split a longer thread at a date and run two turns.
