Getting an email thread into ChatGPT without the noise

An email thread is not one document. It is twenty short documents, most of which quote the previous nineteen, packed in metadata no human has ever read. Get that structure right on the way in and ChatGPT handles threads very well. Hand over the raw file and you spend the conversation explaining what it is looking at.

Why the mail file is the wrong starting point

OpenAI describes what it accepts as "All common file extensions for text files, spreadsheets, presentations, and documents." Mail files are not common documents in that sense, and the two formats fail in different ways.

.msg is Outlook's own binary container. Open one in a text editor and you get noise with a few readable fragments floating in it. Renaming it to .txt changes nothing, because the problem is not the extension, it is that the words are not stored as plain text.

.eml has the opposite problem. It is plain text already, so it generally goes through and the message is genuinely in there. So are the routing headers, the MIME boundaries, and every attachment encoded as base64, which can be thousands of random-looking characters standing in for a logo in somebody's signature.

Neither is worth fighting when the text is one copy away.

Forwarding, and what it is actually good for

There is nothing to forward to. OpenAI does not publish an email address that accepts mail into a chat, so forwarding only moves the thread around your own mailbox.

Forwarding to yourself is still the fastest way to gather a thread. Most clients build the forward as one message with every reply stacked inside it in order, sender and date on each. That is one place to copy from instead of twenty, with the ordering already done.

What to strip

The repetition matters beyond length. Ask what was said about pricing, and a model looking at the same paragraph quoted nine times reads it as the centre of the thread rather than as one remark somebody made in March.

Keep who said what, and when

The one thing never to strip is the attribution. Without it a thread becomes a run of paragraphs with no owner, and almost every useful question about a thread is a question about who.

Reduce each mail header to a single line and keep it:

From: Sam Okafor, 14 March, 09:12

Then check the order. Most clients show newest first, ChatGPT reads top to bottom, and a thread read backwards produces answers in which a decision appears to precede the discussion of it. Say what it is in the first line:

This is an email thread between four people, oldest message first.

1. Expand every message

   Collapsed messages are not included in a copy, and you will not notice the gap until an answer is missing something.

2. Paste into a plain text editor first

   Paste without formatting, then copy again from there. That drops the HTML tables, the tracking pixels and the logos in one move.

3. Delete the repeats

   Signatures, disclaimers, quoted history. This is where most of the size goes.

4. Order it oldest first, attribution intact

   One line of sender and date above each message.

5. Paste it, or save as .txt and upload

   Short threads go straight in the message box. Anything over about 1,500 words becomes an attachment whether you meant it to or not, so save it as a file deliberately.

When the thread is too long

Cleaned up, most threads are small. A year of correspondence with one client is often ten or twenty thousand words, a small fraction of the 2 million tokens a document is allowed.

The limit that actually bites is the message box, and not in the way people expect. Paste more than 10,000 characters and ChatGPT turns the paste into an attachment rather than putting it in the text field, which spends an upload from your daily allowance instead of saving one. That is roughly 1,500 words. Past it, either accept the attachment, use Show in text field to force the text back inline, or send the thread in parts that each stay under the threshold.

Whole mailbox exports are a different problem. Hundreds of threads at once will pass the token cap, and the answer there is to split by thread or by month.

Questions worth asking a thread

Not "summarise this". A thread summary reads like the thread, only shorter.

What was agreed, and where?

What was actually agreed in this thread, and in which message? Quote the line.

Who owes what?

List every commitment anyone made, with the person and the date if one was given. Say "no date given" where there is not one.

What changed?

Did anyone's position change over the course of this thread? Show the before and the after.

What was never answered?

Which direct questions were asked and never answered?

That last one is the strongest, because absences are close to invisible when you read a thread yourself. A question that fell off the end of a busy week is exactly what a reader with no memory of the week is good at spotting.

Before you paste somebody else's mail

Mail is the most personal thing most people upload. It carries other people's addresses, phone numbers, and things they wrote assuming a small audience.

The short version

Copy the thread, do not upload the mail file. Strip signatures, disclaimers and quoted history, keep one sender and date line per message, and put it oldest first. Paste a short thread, upload a long one as .txt, split anything larger than that. Almost every complaint about ChatGPT being vague on an email thread comes from feeding it the same message twenty times with no idea who sent any of them.