Automation: Meeting Notes → Action Items

📖 3 min read resourcesautomations Key Info

Turn a raw meeting transcript into owners, tasks, due dates, and decisions — then create the tasks and post a summary automatically.

On this page 8

OverviewHow it worksYou’ll needBuild it in n8nThe AI promptZapier alternativeCostVariations

Drop in a transcript; get back a clean summary, the decisions made, and a list of action items (with owner + due date) — automatically created as tasks and posted to your team channel.

How it works

flowchart LR
    A[New transcript<br/>Drive / Otter / Fathom] --> B[Send transcript<br/>to the LLM]
    B --> C[LLM returns JSON<br/>summary + decisions + action_items]
    C --> D[Create tasks<br/>Todoist / Linear / Notion]
    C --> E[Post summary<br/>Slack]

You’ll need

• A transcript source — Google Drive file, Otter.ai, Fathom, or a manual paste
• A task tool — Todoist, Linear, Notion, Asana
• An LLM key — Sonnet 4.6 or GPT-5.4 (transcripts are long; mid-tier quality helps extraction accuracy)

Build it in n8n

1. Trigger → Google Drive “New file in folder” (your transcripts folder), or a Webhook you POST the text to.
2. (If a file) Extract from File node → get plain text.
3. HTTP Request → the AI transform node with input = {{ $json.text }}, model claude-sonnet-4-6, max_tokens: 1500, and the system prompt below.
4. Code node → return JSON.parse($json.content[0].text).
5. Split Out node → split action_items into one item per task.
6. Todoist / Linear / Notion node → create a task per item: title {{ $json.task }}, assignee {{ $json.owner }}, due {{ $json.due }}.
7. Slack node → post {{ $json.summary }} + the decisions list to the channel.

The AI prompt

System prompt:

You extract structured outcomes from a meeting transcript.

Return ONLY valid JSON, no prose:
{
  "summary": "<3-4 sentence overview>",
  "decisions": ["<decision 1>", "..."],
  "action_items": [
    {"owner": "<name or 'unassigned'>", "task": "<imperative phrasing>", "due": "<YYYY-MM-DD or 'none'>"}
  ]
}

Rules:
- Only include action items that were actually agreed to, not vague ideas.
- If an owner or due date wasn't stated, use "unassigned" / "none". Do not invent.

Zapier alternative

Trigger (Drive/Otter) → Anthropic → Send Prompt (system prompt above) → Formatter → Parse JSON → Looping by Zapier over action_items → Todoist/Notion → Create Task → Slack → Send Channel Message with the summary.

Cost

A 30-minute meeting ≈ 4–6K input tokens + ~500 output. On Sonnet 4.6 ($3/$15 per 1M) that’s roughly $0.02–0.03 per meeting. Cheap enough to run on every call.

Variations

• Add a calendar trigger so it runs automatically when a recording lands after each meeting.
• Append the summary to a running Notion meeting log keyed by date.
• Add a second LLM pass that drafts the follow-up email to attendees (draft, human-approved).