Turn meeting recordings into searchable knowledge base with transcripts, summaries, and action items

Your team records meetings. You review them never. Six months later, someone asks whether you committed to redesigning the homepage, and nobody can remember the conversation. The recording sits in a folder, unwatched. This pattern repeats in most organisations. Meetings happen, recordings accumulate, and the institutional knowledge trapped inside them evaporates. But what if every recording automatically became a searchable asset? What if action items appeared in your task manager without manual transcription? What if your team could search meeting content across months and actually find what was discussed? The workflow below automates the entire journey from recorded meeting to organised knowledge base. It captures video, generates accurate transcripts, produces summaries, extracts action items, and surfaces them where your team actually works. No manual uploads. No copy-paste transcription. No forgotten commitments. For more on this, see Meeting notes to action items to calendar automation.

The Automated Workflow

This workflow uses MeetGeek to capture and transcribe meetings, then branches into two parallel processing paths: one for searching and analysing content, and one for extracting and tracking action items.

Why n8n for this workflow?

N8n works well here because it handles conditional routing (sending summaries one direction, action items another), manages retries when APIs are slow, and keeps running costs predictable. You could use Zapier or Make, but n8n gives you better control over branching logic without paying per task.

Step 1: Meeting recorded and webhook fires

MeetGeek automatically records your video meetings. When the recording finishes processing, it sends a webhook to your n8n instance with the meeting metadata, transcript, and summary already included.

CopyPOST /webhook/meetgeek-complete
{ "meeting_id": "mtg_abc123xyz", "title": "Q1 roadmap planning", "duration_minutes": 47, "transcript": "Full transcript text here...", "summary": "Key discussion points...", "participants": ["alice@company.com", "bob@company.com"], "recorded_at": "2026-03-15T14:30:00Z"
}

Set up this webhook URL in your MeetGeek account settings under Integrations, then test it by ending a meeting.

Step 2: Parallel processing branches

Once the webhook arrives, n8n splits into two paths: Path A processes the transcript for searchability and long-term storage. Path B extracts action items and assigns them to people. This happens simultaneously, so neither process blocks the other.

Path A: Transcript enrichment and storage

The transcript comes from MeetGeek, but you want it more useful. Send it to Claude Opus 4.6 to add structure: timestamps for key discussion blocks, identified decisions, flagged risks, and highlighted decisions made.

Copy{ "model": "claude-opus-4.6", "max_tokens": 2000, "messages": [ { "role": "user", "content": "Here is a meeting transcript. Add structure by identifying: 1) Decision points with timestamps, 2) Risks or blockers mentioned, 3) Questions left unanswered. Format as JSON.\n\nTranscript:\n{{transcript}}" } ]
}

Store the enriched output in a productivity app like Notion or Obsidian (via their APIs), tagged by date and participant. This becomes your searchable knowledge base.

Path B: Action item extraction and assignment

Send the same transcript to Claude Sonnet 4.6 (faster and cheaper for structured extraction) with a prompt that extracts action items in JSON format.

Copy{ "model": "claude-sonnet-4.6", "max_tokens": 1000, "messages": [ { "role": "user", "content": "Extract all action items from this meeting transcript. For each, identify: who it's assigned to (use the participant names below), what they need to do, the due date if mentioned (or suggest a reasonable deadline based on context). Return as JSON array.\n\nParticipants: {{participants}}\n\nTranscript:\n{{transcript}}" } ]
}

Parse the JSON response and create tasks in your productivity app. Assign them to specific people, set due dates, and link them back to the meeting recording.

Step 3: Task creation in productivity app

Most productivity apps (Notion, Asana, Monday.com) have APIs for creating tasks. Here's the general pattern:

CopyPOST /api/v1/tasks
{ "title": "Redesign homepage hero section", "assignee": "alice@company.com", "due_date": "2026-03-29", "description": "From meeting: Q1 roadmap planning (2026-03-15). Recording: {{recording_url}}", "custom_fields": { "meeting_id": "mtg_abc123xyz", "transcript_link": "{{knowledge_base_url}}" }
}

Step 4: Notification to participants

Once tasks are created, send a Slack or email notification to assigned people. Include a link to the relevant section of the transcript and the full recording.

CopyPOST /slack/webhook
{ "channel": "#project-updates", "text": "New action items from Q1 roadmap planning:", "blocks": [ { "type": "section", "text": { "type": "mrkdwn", "text": "*Alice Chen* , Redesign homepage hero section by Mar 29\n<{{recording_url}}|Watch meeting> | <{{transcript_link}}|Read transcript>" } } ]
}

Putting it together in n8n

In n8n, your workflow looks like this: 1. Webhook Trigger (receives MeetGeek data) 2. Two parallel branches from a "Split in Batches" node 3. Left branch: Claude Opus node for transcript enrichment, then HTTP request to your knowledge base 4. Right branch: Claude Sonnet node for action item extraction, then loop through results creating tasks 5. Final step: Slack notification with summaries Use the "Wait" node if you need to rate-limit API calls, set it to wait 1-2 seconds between Claude calls.

The Manual Alternative

If you prefer more control or want to customise extraction logic heavily, you can skip the automation and do this semi-manually: Download the recording and transcript from MeetGeek, paste them into a Claude conversation, and ask it to extract action items and structure the discussion. Copy the results into your task manager and knowledge base by hand. This takes 10-15 minutes per meeting but gives you direct oversight of what gets extracted. This approach works well if your meetings are highly technical or require judgment calls about priority and ownership. The trade-off is obvious: you lose the speed advantage of full automation.

Pro Tips

Rate limiting and retry logic.

Claude API has rate limits depending on your plan.

In n8n, use the "HTTP Request" node's retry settings: set it to retry 3 times with exponential backoff if you hit a 429 (rate limit) response. This prevents workflows from failing when Claude is under load.

Deduplicate action items.

Sometimes the same task gets mentioned twice in a meeting. Before creating tasks, query your task manager to check if an identical task already exists for that person with the same due date. Skip creation if it does.

Store transcript chunks separately.

If a meeting is over 90 minutes, split the transcript into 30-minute chunks before sending to Claude. This reduces token usage and keeps individual summaries more focused. You can reassemble them for the knowledge base.

Link everything back.

Always include the original meeting ID and a link to the recording in every task, summary, and knowledge base entry. This lets people quickly verify context if there's confusion.

Monitor Claude's output quality.

For the first week, manually spot-check a few extracted action items against the actual recording. Claude sometimes misattributes ownership or invents due dates. If you see systematic errors, adjust your prompt or consider using Claude Haiku 4.5 (cheaper) only for very clear-cut extraction, and Opus for complex meetings.

Cost Breakdown