Thread summarization that's safe to act on: structured asks, quoted evidence, and a checked reply

Step 1: Copy the thread and take out private details

Open the newest email and copy the whole thing, including the older messages quoted below it. Paste it somewhere plain first, like a notes app. Swap names for roles, and remove phone numbers, addresses, account numbers, and anything about a customer you wouldn't put on a postcard. Here's how to remove private details before you paste into ChatGPT.

If it's a work thread, check your company's rules about AI tools first. Some workplaces allow only certain tools, or none.

Step 2: Ask the questions that matter

Paste the cleaned thread and send:

This is an email thread, newest message at the top. I am
[YOUR ROLE]. Tell me:
1. Who is asking me for something, what, and by when.
2. What has already been decided.
3. What is still open or disputed.
4. Anything that changed during the thread (dates, prices,
   names).
For each point, quote the line it comes from.

Telling it which message is newest matters. Threads read backward, and dates that changed midway trip people up.

Step 3: Check the quotes

Skim the quoted lines. If a quote isn't really in the thread, or says something different, don't trust that point. AI tools sometimes fill gaps with something that sounds right. Here's a five-minute way to fact-check an AI answer.

Step 4: Draft the reply, then make it yours

Now ask:

Draft a short reply that answers only the questions
addressed to me. Use plain language. Don't agree to
anything I haven't confirmed. Mark anything you're unsure
about with [CHECK].

Read it against the thread. Fix every [CHECK]. Change any line that commits you to a date or a price you haven't agreed to. Then send it yourself.

Built-in summary buttons

Some email apps now include their own summarize button, depending on your plan and whether your workplace turns it on. It's fine for a quick sense of a thread. For anything you need to answer, the four questions above still do the real work.

Under the hood

Email threads are hard inputs: quoted replies duplicate text, signatures add noise, and reverse-chronological order confuses temporal reasoning. Three controls help:

• Order and role framing. Telling the model "newest first" and who you are fixes many "which date is current?" mistakes.
• Evidence quotes. Requiring a verbatim line per claim turns the summary into something you can audit. A missing or altered quote is a signal the point was invented.
• Scoped generation. Limiting the reply to asks addressed to you, with [CHECK] markers, keeps the model from committing you to things.

Long threads can exceed what a chat window handles well. If the answer misses early messages, paste the thread in two halves and ask for the extraction on each, then merge.

Short close

Copy the thread, take out private details, ask who wants what by when, and make it quote its sources. Then draft the reply, check it, and press send yourself.