Meetings generate decisions and tasks, but notes are often messy and transcripts too long to reread. A well-structured prompt can turn either into a concise record that people actually use. The key is to ask for specific sections and to forbid invention.

The prompt

Below are notes / a transcript from a meeting. Produce a summary with these sections:

1. Purpose of the meeting (one sentence) 2. Decisions made — only decisions explicitly agreed; quote or paraphrase closely 3. Action items — table with columns: task, owner, deadline, source (a short quote showing where it was agreed). If owner or deadline is not stated, write “not specified”. 4. Open questions — issues raised but not resolved 5. Risks or concerns mentioned 6. Next meeting — date and agenda if mentioned

Rules: don’t invent decisions, owners or deadlines. Keep it under 300 words excluding the table. Use neutral language.

Participants: [names and roles, optional]

Notes / transcript: [paste]

Why it works

  • Fixed sections produce consistent, scannable summaries.
  • “Only decisions explicitly agreed” prevents the AI from turning suggestions into decisions.
  • Source quotes let participants verify each action item quickly.
  • “Not specified” exposes missing owners and deadlines — a common reason tasks get dropped.
  • A word limit keeps the summary readable.

Handling long transcripts

Research such as Lost in the Middle has found that language models can be less reliable at using information placed in the middle of very long inputs. For long meetings:

  • split the transcript into parts, summarize each with the same structure, then merge;
  • ask the model to check specifically for action items in each part;
  • review the middle of the meeting carefully, where details are easiest to miss.

Variations

Executive version: “Summarize in five bullet points for a senior manager who wasn’t there: decisions, impact, what’s needed from them.”

Follow-up email: “Draft a friendly follow-up email to participants with the decisions and action items, asking them to confirm or correct within two days.”

Project tracker format: “Output action items as CSV with columns task, owner, due_date, status=open” for importing into a task tool.

Multilingual meetings: “The transcript mixes English and Italian. Write the summary in English, translating quotes.”

Privacy and accuracy

  • Get participants’ consent before recording and transcribing meetings, and follow your organization’s rules.
  • Use approved tools for confidential discussions.
  • Always have someone who attended review the summary before sharing — speech recognition errors and misattributed statements are common.

Example output (abbreviated)

Purpose: agree the launch date for the new onboarding flow. Decisions: launch moved to 14 March; the email sequence will be shortened to three messages. Action items: update the launch plan — Marta — 1 March; draft the three emails — Luca — not specified. Open questions: whether the help center articles need translation before launch.

Adapting the prompt to your team

  • For recurring meetings, add: “Compare with last week’s action items below and mark each as done, in progress or not mentioned.”
  • For client meetings, add a section for “commitments we made to the client” and “information requested from the client”.
  • For project kickoffs, add “scope”, “assumptions” and “success criteria” sections.
  • For incident reviews, use sections for timeline, impact, root causes, and follow-up actions, and keep a blameless tone.

Save your adapted version as a template in your notes tool or assistant so the whole team produces summaries in the same format. Consistency makes it easier to scan past meetings and track whether action items were completed.

Common failure modes to check

  • Assigning tasks to the wrong person when several people speak about the same topic.
  • Turning discussion into decisions, for example treating “we could try” as “we decided”.
  • Dropping numbers and dates mentioned in passing.
  • Missing tasks agreed near the end of a meeting, when people rush.

What to watch next

  • Meeting tools generating action items automatically.
  • Integration of summaries with task and project tools.
  • Workplace rules on recording and AI transcription.
  • Better handling of long transcripts by models.

Key takeaways

  • Ask for fixed sections: purpose, decisions, action items, open questions, risks and next steps.
  • Forbid invented decisions, owners and deadlines; use “not specified”.
  • Source quotes make action items easy to verify.
  • Split long transcripts to avoid missing details in the middle.
  • Get consent for recording and have an attendee review summaries.

Frequently asked questions

Can AI join meetings and take notes automatically?

Many meeting tools offer AI note-taking. Check consent requirements and your organization’s policies.

How accurate are AI meeting summaries?

Usually good for main points, but they can miss details or misattribute statements. Review before sharing.

What if the notes are very rough?

The prompt still helps; it will flag missing owners and deadlines, which you can fill in.