Prepare the meeting and prompts

Before asking AI for an agenda, write a brief with five parts:

1. Purpose: Why does this meeting exist?
2. Outcome: What should be true when it ends?
3. Participants: Who is attending, and what authority or knowledge do they bring?
4. Context: Which facts, documents, constraints, and prior decisions matter?
5. Boundaries: What is out of scope, confidential, undecided, or owned elsewhere?

Keep the brief factual. “The team is behind because engineering underestimated everything” is an accusation, not neutral context. Prefer “Milestone B is eleven days later than the plan dated 4 June; the cause has not been agreed.” Separating facts from interpretations helps the model avoid turning one person’s view into the meeting’s premise.

Do not paste every related document. Include the smallest set that supports the intended outcome. Label documents with dates and status: “approved budget, 3 May,” “draft launch plan, 12 June,” or “superseded timeline.” AI cannot reliably infer which of several contradictory files governs the discussion.

A useful preparation prompt assigns a role, supplies context, defines output, and prohibits invention:

You are helping prepare a 30-minute launch-readiness meeting. Using only the brief below, propose an agenda with time boxes. The required outcome is a go/no-go decision for 15 September plus owners for unresolved blockers. Separate information-sharing, discussion, and decision points. Identify missing context as questions; do not fill gaps. End with a two-minute recap of decisions and actions.

The phrase “using only the brief” does not guarantee perfect grounding, but it states the rule you will review against. “Identify missing context” gives the assistant a safe alternative to guessing. Time boxes expose an agenda that cannot fit. Explicit decision points prevent status updates from consuming the whole session.

Review the result as a facilitator. Check that the people with decision authority are present, pre-reading is realistic, and the hardest item receives enough time. Remove decorative sections that do not change the outcome.

An agenda prompt and a note-taking prompt do different jobs. The first designs the conversation. The second defines what should be captured from it. Write capture instructions before the call:

• record decisions only when the group clearly commits;
• capture action, owner, and due date as separate fields;
• label proposals, concerns, and open questions distinctly;
• preserve important numbers and dates exactly;
• mark inaudible or ambiguous passages instead of guessing;
• include the evidence or transcript location behind consequential claims;
• exclude side conversation that does not affect the outcome.

Avoid “take detailed notes.” Detail is not a useful specification. It can produce a long transcript-shaped document with no hierarchy. Define the readers instead: an attendee who needs a reminder, an absent stakeholder who needs context, or an operator who must execute the actions. Different readers need different outputs.

Meeting capture often fails on product names, acronyms, and people’s names. Give the tool a short vocabulary list such as “Aster (project), Nia Chen (security), SOC 2 (compliance), EUR (currency).” Include likely spelling variants where useful.

This list helps recognition, but it can also bias transcription. If “Aster” is in the list, the system may replace a similar-sounding ordinary word with the project name. Treat the vocabulary list as a hint, then verify its appearances. Never use it to predeclare what participants will conclude.

Ask AI to challenge the brief before generating the agenda:

• Which decision-maker appears to be missing?
• Which requested decision lacks a criterion?
• Which claims need a source or current measurement?
• Which agenda items could be handled asynchronously?
• Which terms could mean different things to different participants?

Review each question rather than accepting the model’s diagnosis. A model may recommend inviting more people because it does not understand delegated authority. Its questions are prompts for human judgment, not proof.

Five minutes before sending the agenda, check:

• The outcome is written as an observable result.
• Decision rights are named.
• Source documents have dates and status.
• Sensitive material has an appropriate audience.
• The agenda fits the available time.
• Capture instructions distinguish facts, proposals, decisions, and actions.
• The recording or transcription plan will be disclosed before capture begins.
• A non-recorded alternative exists if someone cannot consent.

Preparation should reduce meeting time, not add a bureaucratic ritual. Save reusable templates for recurring one-to-ones, project reviews, interviews, and incident calls, but update the context each time. A stale template can preserve an obsolete decision or expose yesterday’s attendee list.