Using AI to Prepare for a Meeting

A single pre-read is the wrong output. What you actually use in a meeting is three different things, produced by three different requests, and each one is only trustworthy if every line names where it came from.

Three artefacts, not one

Asking for “a briefing for my meeting” produces a document that is too long to hold in your head and too shallow to answer anything. Split it.

The separation matters because the failure modes differ. A wrong fact in the briefing embarrasses you. A missing question wastes the meeting. A missing objection loses the argument. Producing them together guarantees that the two harder ones are done badly.

Gathering the material

Everything here depends on giving the model the actual material. A briefing written from the model’s general knowledge about your counterparty is a briefing about a plausible company that does not exist. The rule for the whole page: no line without a source, and the sources are things you supplied.

• The email thread. Paste it whole, including the parts you think are irrelevant. Positions and grievances are usually in the middle of a thread rather than at the end.
• The last meeting’s notes or minutes. The single most useful document, because commitments made and not kept are the substance of most difficult meetings.
• The document under discussion. The proposal, the contract, the report, the numbers.
• Anything public you have gathered yourself. Paste the text of the annual report page or the press release, not a request to go and find it. If the tool does have web access, every fact it returns from the web still needs the check in fact-checking what an AI told you.

The briefing

From the material below only, write a one-page briefing for a
meeting on Thursday with [counterparty] about [subject].

Structure:
1. WHERE THINGS STAND - four sentences maximum. The current state,
   not the history.
2. FACTS I MUST NOT GET WRONG - a list. Names and roles, figures with
   their date, deadlines, anything already agreed in writing.
3. OPEN ITEMS - things raised and not resolved, with who raised them
   and when.
4. WHAT CHANGED SINCE LAST TIME - only if the material shows it.

Every line ends with its source in square brackets: [thread, 14 July]
or [minutes, item 4] or [proposal, p.3].

If a fact is not in the material, it does not go in. Where you think
something important is missing, list it under "NOT IN THE MATERIAL"
at the end.

The bracketed sources are not decoration. They make the briefing checkable in ninety seconds and they make one specific failure visible: a line with a source you cannot find is a line that was invented. The “not in the material” section is often the most valuable part, because it tells you what to go and find before Thursday.

The question list

Good questions come from gaps, and gaps are visible in the material if you ask for them directly rather than asking for questions.

From the same material, produce a question list.

First, identify the decisions this meeting could reasonably reach.

Then for each decision, give the questions I need answered to make
it. For each question:
- the question, in the words I would say it
- why I am asking it (what I do differently depending on the answer)
- what a non-answer looks like, so I notice one

Order by what I cannot decide without. Maximum ten. Mark the three
I should ask even if the meeting runs short.

The “what a non-answer looks like” line is the one worth having. In a difficult meeting the common outcome is not a refusal but a fluent reply that does not contain the information, and it is very hard to notice in real time unless you have written down in advance what you are listening for.

The objection list

Ask for the strongest version of the other side’s position, not a list of things they might say. Weak objections are useless to rehearse against.

You are [counterparty]'s operations director. You have read the
material below and you do not want to agree to [the proposal].

Write the five strongest objections you would raise, in your own
words, strongest first. For each: what you actually want instead.

Then, separately, as me: a response to each objection, using only
facts from the material, and marking any response that depends on
something I have not yet verified.

Do not soften the objections. If one of them is unanswerable from
the material, say so rather than inventing an answer.

The instruction not to soften matters. Left alone, the default output is agreeable, and an agreeable simulation of your counterparty is worse than no preparation because it builds false confidence. There is more on that failure, and how to prompt around it, in preparing for a difficult conversation.

If you are chairing it

Chairing is a different preparation problem: your job is not to win an argument but to make sure the meeting reaches a decision, and most meetings fail at that for structural reasons that are visible in advance.

1. Write the decisions, not the topics. An agenda item called “budget” produces a discussion; one called “decide whether to approve the £40k spend, or name what we need to decide it next time” produces an outcome. Ask for each agenda item to be rewritten as the decision it is meant to reach, and for the items that are not decisions to be marked as information only.
2. Time-box against reality. Total the times and compare them with the length of the meeting. Almost every agenda fails this check, and it is easier to cut an item beforehand than to abandon one at five to the hour.
3. Identify the items that need a pre-read and what that pre-read must contain for the decision to be possible. An item with a missing input is an item that will be deferred; better to know on Monday.
4. Ask what could derail it. Which item will overrun, which two attendees disagree, which decision somebody absent will want to reopen. You will know most of this; having it listed is what makes you willing to interrupt.

Every one of those is a transformation of material you already have, which is the pattern that makes this work at all. Nothing here asks the model to know anything about your organisation.

The twenty-minute version

1. Paste the thread and the last minutes. Ask for the briefing prompt above. Five minutes.
2. Read it and fix the errors yourself. You will find one or two, and correcting them in your own head is most of the preparation. Five minutes.
3. Run the question prompt against the corrected briefing rather than the raw material. Three minutes.
4. Run the objection prompt. Read the objections, ignore the responses, write your own for the two that landed. Seven minutes.

What you take into the room is one page of facts and one page with ten questions and five objections on it. Not the transcript of a conversation with a chatbot.

If the meeting is one you have every month with the same people, the briefing from last time is your material for this time, and the whole preparation collapses to about eight minutes. The compounding is the reason to keep the outputs in a file rather than in a chat history.

Afterwards: notes, actions and the record

The half-hour after a meeting is where more value is available than in the preparation, and it is almost always skipped. Write your own rough notes first — five minutes, ugly, from memory — and use the model to structure them rather than to remember for you.

Here are my rough notes from a meeting. Turn them into two things.

1. A RECORD, for me: what was said, in the order it was said,
   in plain sentences. Use only my notes. Where my notes are
   ambiguous, write [UNCLEAR: ...] rather than resolving it.

2. AN ACTIONS EMAIL, to send to the other side, containing only:
   - what we agreed, one line each
   - who does what, by when
   - what was explicitly not agreed or was deferred
   Anything I did not write down does not go in.

At the end, list QUESTIONS FOR ME: things my notes imply were
decided but do not say clearly.

[rough notes]

The actions email is the artefact that prevents the commonest failure of difficult meetings, which is that both sides leave believing different things were agreed. Sending it the same day, with the non-agreements listed as plainly as the agreements, converts a conversation into a record at almost no cost.

What it cannot prepare you for

It does not know the people. Everything about who defers to whom, who has been overruled, who is leaving in March and who has an unrelated grudge is outside the material and is usually what decides the meeting. No amount of prompting recovers it.

It will produce a confident briefing from thin material. If you paste two emails, you get a page, and the page will not look thinner than one built from twenty documents. The “not in the material” section is the only defence, which is why it is in the prompt.

And it cannot tell you what you are wrong about, because it has only your side’s documents. If the material is entirely your own emails, the objection list is your imagination of the other side, structured. That is worth having. It is not intelligence.