AI Prompt to Summarise a Long Negotiation Email Thread

The ai prompt to summarize a negotiation email thread that actually works is narrower than a general thread summary — it outputs a position ledger rather than a narrative recap. A narrative tells you roughly what happened. A ledger tells you who opened where, exactly what each side conceded and when, which terms are now locked, which are still live, and which deadlines are attached to what. That is the format a deal thread needs, because the expensive mistake in a negotiation is not misremembering the general direction — it is confusing a proposal for a commitment, or missing the concession that shifted the anchor three messages back.

General-purpose summary prompts tend to produce a paragraph of even-handed prose that treats every message as roughly equal weight. A negotiation thread is not like that. Certain sentences are load-bearing — the first number, the concession that moved the other side, the word 'agreed' that locked a term — and they need to be extracted, attributed, and quoted verbatim rather than folded into a smooth narrative that softens '$49,500' into 'approximately $50,000' or merges a conditional offer with a settled deal.

This post gives you the prompt, the preparation steps, a table of how major platforms handle it differently, and the failure modes that quietly corrupt negotiation summaries. At the end, if you negotiate regularly across multiple threads, there is a faster path that does not require copying anything out of your inbox each time.

What do you need before running the prompt?#

Before a single prompt is worth running, the input has to be right. Paste the thread oldest-first — the first offer or opening message at the top, the latest reply at the bottom. Email clients display threads newest-first by default, so you are reversing the order. If you can only get the thread newest-first, add one line at the top of your prompt telling the model the thread is in reverse chronological order; it will adjust, but oldest-first produces fewer attribution errors on the concession timeline.

Keep sender names and timestamps. The ledger format requires the model to attribute each concession to a named party on a specific turn. Strip those, and the best it can do is note that a figure moved — it cannot tell you who moved it or when. Also keep any defined terms, version numbers, and capitalised nouns (words like 'Services', 'Deliverables', or 'Effective Date') because these are the exact strings the verbatim-quoting rule asks the model to reproduce.

Delete what does not carry meaning: the repeated quoted blocks that email clients insert under every reply, legal disclaimers, unsubscribe footers, and signature blocks. In a negotiation thread that has run for weeks, these can easily triple the character count without adding a single fact. Trimming them gives the model more context budget for the actual conversation and reduces the chance that a boilerplate confidentiality notice gets treated as a negotiating position.

What does the position-ledger prompt look like?#

Copy this prompt and paste your prepared thread beneath it. The five-section structure is the part that separates this from a generic summary. Each rule about verbatim quoting is doing real work: a model told to summarise will round numbers, soften qualifiers, and conflate a proposal with an acceptance. The instruction to forbid paraphrase stops all three of those errors at once.

How do you run the prompt on a real thread?#

1. 1

   Export the full thread

   Open the thread in your email client and expand all collapsed messages. In Gmail, use the expand icon at the top right of the conversation view so every message is visible, not just the latest. In Outlook, choose 'Show all messages' in the thread view. Select all, copy as plain text. Do not copy only the top message — the model will summarise one reply and present it as the whole thread.

2. 2

   Reverse the order if necessary

   Paste the text into a plain-text editor and check whether the oldest message is at the top. If it is newest-first, either scroll to the bottom of the original thread and copy from there upward, or add the line 'Note: this thread is pasted newest-first; the last message is the most recent' at the very top of your prompt input so the model does not mistake the final reply for the opening position.

3. 3

   Strip boilerplate without stripping facts

   Delete repeated quoted blocks, legal disclaimers, confidentiality footers, and signature lines. Do not delete defined terms, capitalised nouns, or any line that includes a number, a date, or a percentage — those are the raw material for the verbatim quotes the prompt requires. A quick pass to remove obvious noise takes one minute and measurably improves output quality.

4. 4

   Paste the prompt, then the thread

   Open your chosen AI platform, paste the full position-ledger prompt above, then paste the cleaned thread immediately below it. Do not leave a blank line or separator between the prompt and the thread that might confuse the model about where the instructions end and the data begins. Run the prompt on the full thread in one pass for threads of ordinary length.

5. 5

   Spot-check the output against the source

   For every settled term, open the original thread and confirm the quoted phrase matches the source exactly. For the concessions list, verify the attribution and the chronological order. The ledger is only as reliable as this check — a fabricated date or a misattributed concession looks identical in formatting to a correct one, and both feel equally authoritative when you read them later.

Does the prompt work the same across AI platforms?#

The prompt is model-agnostic — the same five sections work in any capable large language model. The platforms differ in how you get the thread in front of the model and what constraints apply. Verify current capabilities against each vendor's live documentation before relying on any specific feature, as these surfaces change frequently.

What do you do when the output is wrong?#

Three failure modes appear most often on negotiation threads, and each has a specific fix rather than a general 'try again.'

The most common is a concession logged as a settled term. The model reads 'We can come down to $45,000' as agreement because in ordinary prose it looks like one. Fix: the rule 'A settled term requires explicit acceptance from both parties; a one-sided movement is a concession, not a settled term' needs to be in the prompt — it is already included in the version above, but if you are running a shortened version of the prompt, this is the line to restore first.

The second is paraphrased numbers despite the verbatim instruction. This usually happens when the thread is very long and the instruction gets diluted across a large input. Fix: chunk the thread into chronological halves, run the ledger prompt on each chunk separately, then run a final merge pass. Paste the two partial ledgers and instruct the model to reconcile them into one — noting any item that was open in the first chunk and settled in the second. Each chunk is short enough that the verbatim rule holds throughout.

The third is a section filled with invented content — the model hallucinates a deadline that was implied but never stated, or attributes a concession to the wrong party. Fix: the instruction 'If a section is empty, write None found — do not infer or estimate' needs to be explicit. Most hallucination in negotiation summaries comes from the model trying to be helpful by filling a slot the format created. The explicit None found permission removes the pressure to fabricate.

A faster way to track every negotiation thread#

We build AI Emaily — an AI-native email client — and the position-ledger summary is one of the things it handles without any copy-paste loop. When you open a deal thread, AI Emaily reads it inside your inbox with the full message history, sender attribution, and timestamps already intact. You do not paste, reorder, or strip boilerplate manually; the client has the context the prompt above asks you to reconstruct by hand.

The resulting ledger stays attached to the thread it came from rather than stranded in a separate chat window, so you always have a live view of the negotiation alongside the actual correspondence. In Copilot mode, every proposed action — a drafted reply, an extracted commitment, a flagged deadline — waits for your explicit approval before anything leaves your outbox, with undo and a full audit trail on every step. Your negotiations are treated as confidential and are never used to train models.

If you run deal threads regularly across Gmail, Outlook, iCloud, or any IMAP account, see how it works at aiemaily.com. Details on the 7-day free trial and plan options are on the pricing page at aiemaily.com/pricing.