There's a specific kind of dread that comes with opening an email thread that's been running for three weeks.
You know the one. Twenty-two messages. Four different people. Three separate side conversations that got merged in by accident. A subject line that says "Re: Re: Re: FWD: Quick question" and hasn't been accurate since the first message.
And somewhere in there—buried between a scheduling back-and-forth and someone's out-of-office reply—is an actual decision that somebody needs to act on.
I used to handle these by reading the whole thing top to bottom, twice, taking notes the second time. It took twenty minutes and I still usually missed something.
Now I do it in about six minutes. Here's the workflow.
The Problem With Long Threads
Email threads are chronological, but decisions aren't.
The information you need is scattered by time, not by importance. Message 3 has the original request. Message 14 has the constraint that changes everything. Message 19 has the actual decision. Message 21 has someone quietly walking it back.
Reading in order means you're building the whole picture in your head as you go, and by message 18 you've forgotten what was in message 4. This is not a personal failing. It's just how human memory works when you're reading a conversation that wasn't designed to be read.
ChatGPT is good at this because it doesn't read chronologically. It reads everything at once and can reorganize by meaning.
What I Do Before I Paste Anything
This comes first, every time, and I want to say it before the method because it's the part that matters most.
Email threads are the single most dangerous thing to paste into ChatGPT. Not because ChatGPT is malicious, but because threads are dense with identifying information. Names, email addresses, client details, internal project names, sometimes attachments referenced with sensitive titles.
So here's my process.
Step one: Check the policy. I know what my company allows. If you don't know what yours allows, find out before you do anything else.
Step two: Strip it down. I open the thread in one window and a blank document in another. I copy the content—what was said, what was decided, what's unresolved—and I replace every name with a role. "Person A," "the client," "the vendor." I remove email addresses, phone numbers, account numbers, and anything marked confidential.
Step three: Ask myself whether it's worth it. If the thread can't be anonymized without becoming useless, I don't paste it. I read it the slow way. Some threads are too sensitive to run through any external tool, and that's fine. The workflow is a convenience, not a right.
Only after those three steps do I paste anything.
The Prompt
Here's what I run.
"Below is an email thread. I need a decision summary, not a summary of the conversation.
Produce four sections:
The Decision — what was actually decided, or 'NO DECISION MADE' if nothing was settled.
Who Owns What — action items with owners, or 'UNASSIGNED' if no owner was named.
Unresolved — questions that were raised but never answered.
Contradictions — anywhere the thread says two different things, or where someone changed their position.
Rules: Do not invent names, dates, or details. Do not summarize small talk, greetings, or scheduling logistics. If something is ambiguous, flag it rather than resolving it. Quote the specific message where the decision was made, using the message number."
That last instruction—quote the message—is the one that turned this from a nice idea into something I actually trust.
Why I Ask for the Citation
Because it lets me verify.
If ChatGPT says a decision was made in message 14, I can jump to message 14 and check. That takes ten seconds. Without the citation, I'd have to either trust the summary completely or re-read the whole thread—and one of those options defeats the purpose.
This is the same principle as the "do not invent" rule I use everywhere else. I'm not asking ChatGPT to be correct. I'm asking it to show its work so I can check.
Why "Contradictions" Is Its Own Section
This one earns its place constantly.
Long threads almost always contain a contradiction somewhere. Someone agrees to a timeline in message 8 and quietly adjusts it in message 16. A scope gets widened and then partially narrowed. Two people describe the same requirement differently and nobody notices.
When I used to read threads manually, I'd catch maybe half of these. Now they get flagged explicitly.
I've had at least three cases where this section caught something that would have caused a real problem—someone had agreed to something in a side message that contradicted the main thread, and nobody had reconciled it.
A Real Example
Here's a trimmed version of an actual thread I processed. Names and specifics changed, structure accurate.
Messages 1–4: A client asks whether a delivery date can move up by a week. Two people discuss feasibility.
Messages 5–9: Back-and-forth about whether the warehouse can accommodate it. Someone says it's possible "if we push the other order."
Messages 10–13: Scheduling logistics and a tangent about a separate project.
Message 14: A manager says, "Let's go with the original date and revisit if anything changes."
Messages 15–18: Someone asks a follow-up question that nobody answers. A different person mentions a concern about the other order getting delayed.
Message 19: A team member writes, "Agreed, we'll keep the original date."
Messages 20–22: Out-of-office replies and a thank-you.
Here's what the prompt produced.
The Decision
The delivery date stays as originally scheduled. The request to move it up was not approved. (Decision made in message 14, confirmed in message 19.)
Who Owns What
Revisit the date if anything changes — Owner: UNASSIGNED
Address the concern about the other order being delayed — Owner: UNASSIGNED
Unresolved
A follow-up question in message 15 was never answered. The question was whether the other order would still ship on time if the date had moved.
Contradictions
Messages 5–9 imply the earlier date is feasible. Message 14 says to keep the original date. The reason for the reversal is not stated in the thread.
Look at that last item.
Two people spent messages 5 through 9 establishing that moving the date up was possible. Then in message 14, a manager reversed it, without explanation. If I'd read this thread the normal way, I might have concluded the date was moving—because that's what most of the thread is about.
The contradiction flag caught it. That's the entire value of this workflow in one example.
The Test Card
Task: Extract the actual decision, owners, and open questions from a long email thread.
Prompt or workflow: Anonymize the thread, then run the four-section decision prompt with citation and contradiction requirements.
Starting conditions: A thread of roughly 20 messages with multiple participants, side conversations, and at least one scheduling tangent.
Time before ChatGPT: About 20 minutes of reading, plus note-taking, plus a real chance of missing a reversal or an unanswered question.
Time after ChatGPT: About 6 minutes total—3 minutes anonymizing, 90 seconds generating, 90 seconds verifying citations.
Editing required: Minimal on the output. The main work is the anonymization step, which is non-negotiable.
What went wrong: Early versions summarized the conversation instead of extracting the decision, and treated discussion as agreement. Fixed by asking specifically for decisions and adding the contradiction section.
Privacy notes: Every name replaced with a role. Email addresses, phone numbers, account numbers, client identifiers, and confidential project names removed before pasting. Company policy checked first. Threads that can't be meaningfully anonymized are read manually.
Who should not use this method: Anyone in a workplace that prohibits external AI tools. Anyone working with threads they can't strip of identifying detail. And anyone who will skip the citation check—the citations are what make the summary trustworthy, and skipping them means you're trusting an unverified summary of a thread you didn't read.
What This Changed
A few things, honestly.
I catch reversals now. The contradiction section has flagged something in maybe a third of the threads I've processed. Not all of them were important. Some were. The important ones would have embarrassed me.
I stopped re-reading. That's the real time savings. I used to read threads twice—once to understand, once to extract. Now I read once to anonymize, and the extraction is done for me.
I write better follow-ups. When the summary flags an unanswered question, I can send a short email that resolves it instead of letting it sit. Three sentences instead of a fifteen-minute re-read of the thread.
I stopped pretending I remembered things. Twenty-two messages is more than my brain holds reliably. Admitting that and building a process around it has been more useful than trying to be the person who remembers everything.
The Bottom Line
Long email threads aren't hard because they're long. They're hard because they're organized by time and you need them organized by meaning.
That's a mechanical problem, and mechanical problems are exactly what ChatGPT is good at. You give it the mess, it gives you the structure, and it shows you where the thread contradicts itself.
But the anonymization step isn't optional, the citation check isn't optional, and some threads should never be pasted anywhere. Those are the rules, and I don't bend them.
I tried it at my desk so you don't have to. Useful beats impressive—and knowing what was actually decided is about as useful as it gets.
