AI Meeting-Summary & Action Extraction Prompt Pack

Purpose

This prompt pack helps you extract the few things that actually matter from meetings: concise summaries, explicit decisions, clear action items (who's doing what, by when), and simple prioritization. Use these prompts as starting points — tune them for your meeting style, team terms, and required level of formality. Always run a quick human verification step before publishing or acting on extracted items.

When to use this pack

• After a recorded meeting or live notes are available.
• To prepare meeting follow-ups, status updates, or change logs.
• To create a shared, auditable list of decisions and owners.

How to prompt: general guidance

Provide the model with context (meeting title, date, attendee list) and the raw transcript or notes. Ask for a structured output (JSON or bullet lists) to make downstream processing and handoffs reliable. Use short explicit system-style instructions to fix output shape and reduce hallucination.

Core prompts and templates

1) Short meeting summary (concise)

Prompt (short):

"Read the meeting transcript below. Provide a single-paragraph summary (2–4 sentences) that captures the meeting purpose, main outcome, and next step(s). Keep it factual and neutral."

2) Action-item extraction (structured)

Goal: reliably extract actionable tasks with owner, due date, priority, and any clarifying notes.

Prompt (template):

"From the transcript below, list all explicit action items. For each action item provide: id, task (one sentence), owner (name or role), due_date (ISO date if stated, otherwise "unspecified"), priority (high/medium/low if mentioned or infer based on urgency), source_timestamp (if available), and notes (short clarifying text). Output as a JSON array. If an item is ambiguous about owner or date, set owner/date to "unspecified" and add a note explaining the ambiguity."

Example JSON output (model output shape):

[
  {
    "id": "A1",
    "task": "Prepare draft budget for Q4",
    "owner": "Maria Gomez",
    "due_date": "2026-10-15",
    "priority": "high",
    "source_timestamp": "00:12:34",
    "notes": "Include projected hiring costs; Maria to confirm headcount assumptions."
  }
]

3) Decision extraction

Prompt (template):

"From the transcript below, extract formal decisions made, including: decision_id, decision_text (one sentence), rationale (brief), decision_maker(s) (person or group), effective_date (if stated), and implications (short list). Output as JSON array."

4) Prioritization prompt (quick)

Use when you want the model to suggest an ordering for action items.

"Given the list of action items below, assign a priority (1 highest — 5 lowest) using this simple rubric: 1 if blocking work or regulatory, 2 if high impact within 2 weeks, 3 if medium impact within 1 month, 4 if low impact or longer timeline, 5 if optional/parking. Return the list as JSON with id and priority_score and a one-line justification for each."

5) One-step meeting recap for email

Prompt (template):

"Write a short meeting follow-up email subject and body that includes: one-line summary, 3–5 action items with owners and dates, and a call to confirm or correct within 48 hours. Keep tone professional and helpful."

Tuning tips

• Set a firm output schema (JSON) to avoid freeform text that is hard to action.
• For production pipelines, use few-shot examples (2–3 examples) that match your preferred JSON keys and phrasing.
• If your transcripts contain many speakers with similar names, provide an attendee list mapping short names to full names to reduce owner ambiguity.
• Use lower creativity/temperature for extraction tasks to reduce hallucination.

Human-review handoff checklist (must-run before distribution)

Mark items on the checklist before publishing or acting on extracted outputs:

1. Verify owner names against attendee list — fix misspellings and role mismatches.
2. Confirm deadlines — if a date is unspecified, assign an owner to propose one.
3. Resolve ambiguous tasks — add clarifying notes or follow-up questions where the model marked "unspecified."
4. Check decisions for required approvals — escalate items that need formal sign-off.
5. Validate priority assignments with the meeting lead for alignment.
6. Confirm that any regulatory, compliance, or safety items are routed to responsible functions.
7. Publish with a clear change log and source reference (meeting id, transcript filename, timestamp).

Safety & quality reminders

• AI outputs can omit context or invent details. Do not treat extracts as authoritative without human verification.
• Watch for inferred owner assignments — prefer explicit confirmation from participants.
• Do not publish confidential or legally sensitive summaries without review by appropriate stakeholders.

Integration and agent handoff suggestions

To make this pack operational in a workflow or agent pipeline:

• Standardize the model output JSON keys (id, task, owner, due_date, priority, source_timestamp, notes) so downstream tools can ingest reliably.
• Consider storing extra context (meeting_id, transcript_uri, attendee_map) in the same record for traceability.
• When possible, have an assigned human reviewer confirm or edit the extracted JSON before it is posted to shared channels or task trackers.

Quick examples (input → output)

Input (excerpt): "...John: Maria, can you draft the Q4 budget? Maria: Yes, I'll have it by Oct 15. ..."

Action-item JSON (expected):

[{ "id": "A1", "task": "Draft Q4 budget", "owner": "Maria", "due_date": "2026-10-15", "priority": "high", "source_timestamp": "00:12:34", "notes": "John requested including hiring costs." }]

Common mistakes and how to avoid them

• Missing owners: provide an attendee map or instruct model to mark "unspecified" rather than guessing.
• Vague tasks: ask model to reduce tasks to one clear imperative sentence starting with a verb.
• Inconsistent dates: normalize dates to ISO format in the prompt and request "unspecified" when not clearly stated.

Customization examples

Teams may want to add fields like "estimate_hours", "related_project", or "compliance_flag". Add these to the JSON schema in the prompt and provide a one-line example for each.

Wrap-up

Use these templates as reusable patterns. Start with conservative extraction (mark unknowns instead of guessing), require a human check from the meeting owner, and iterate by adding examples from your team's meetings to improve the model's consistency.