Meeting Transcriber

This skill guides transformation of raw meeting transcripts into structured, actionable summaries. The user provides unstructured text from a meeting (transcript, voice-to-text output, or rough notes) and optionally specifies focus areas (decisions only, action items, full summary). Apply the workflow below, then return a structured summary with clear ownership and deadlines.

🎯 When to Use

• User pastes a meeting transcript, Zoom/Teams/Google Meet export, or voice-to-text output
• User shares rough meeting notes and requests "clean this up" or "what did we decide?"
• User asks to extract action items, decisions, or key discussion points from a conversation
• Transcript contains multiple speakers, timestamps, or disorganized dialogue
• Do NOT use when the user wants to generate a meeting agenda from scratch, schedule meetings, or transcribe raw audio/video files — refer them to appropriate transcription or scheduling tools
• Do NOT use for non-meeting conversations (customer support logs, job interviews, legal depositions, therapy sessions) — these require domain-specific workflows
• Do NOT use when the input is already a clean, structured summary — no processing needed

🧠 Core Workflow

Step 1 — Ingest & Parse Detect the input format: raw transcript (speaker labels + timestamps), voice-to-text (no speaker separation), or bullet notes. Identify:

• Number of speakers (detect patterns like "Speaker 1:", "[John]:", etc.)
• Presence of timestamps and their format
• Meeting metadata if available (date, title, attendees list) If speaker labels are missing or inconsistent, normalize to Speaker 1, Speaker 2, etc. and flag this limitation in warnings.

Step 2 — Segment by Topic Group dialogue into thematic clusters using keyword detection and conversation flow:

• Detect topic shifts via phrases like "moving on to", "next item", "let's discuss"
• Cluster related exchanges even if interrupted by tangents
• Label each segment with a descriptive heading (e.g., "Q3 Budget Review", "API Redesign Discussion") If the transcript lacks clear topic boundaries, segment by time intervals (every 5-10 min) and infer topics post-hoc from content analysis.

Step 3 — Extract Structured Elements For each segment, identify and categorize:

1. Decisions — Look for explicit agreement markers:
   • "Let's do X", "We agreed on Y", "Decision: Z", "Okay, we'll go with..."
   • Capture the decision, who proposed it, who approved it, and any dissenting opinions
   • Flag ambiguous outcomes ("we'll think about it") as ⚠️ pending rather than decisions
2. Action Items — Extract tasks with the following schema:
   • Owner: Who is responsible (detect "I'll...", "[Name] will...", "assign this to...")
   • Task: What needs to be done (verb-first description)
   • Deadline: When it's due (explicit dates, "by EOW", "before next meeting")
   • Priority: Infer from urgency language ("ASAP", "blocker", "nice-to-have") If owner/deadline is missing, flag as ⚠️ incomplete action item and suggest follow-up
3. Key Discussion Points — Summarize substantive exchanges:
   • Capture arguments for/against proposals with attribution
   • Note data points, metrics, or evidence cited
   • Flag unresolved debates or open questions requiring future discussion
4. Risks & Blockers — Identify impediments mentioned:
   • Dependencies on other teams, technical constraints, resource gaps
   • Capture severity and proposed mitigation if discussed

Step 4 — Validate & Cross-Check Review extracted elements against the original transcript:

• Verify action item owners are actual participants (not misattributed)
• Confirm deadlines are realistic and explicitly stated (not inferred unless obvious)
• Check that decisions reflect consensus, not unilateral statements
• Flag any contradictions (e.g., two conflicting decisions in same meeting) If confidence is low on any extraction, include it with a ⚠️ low-confidence warning rather than silently omitting or fabricating details.

Step 5 — Return Output Provide structured JSON report followed by optional human-readable executive summary. Always include attendee inference, meeting duration (if timestamps available), and completeness score (% of action items with clear owner + deadline).

📋 Output Format

Return two sections:

Section 1 — Meeting Summary Report

{
  "status": "success | partial | error",
  "metadata": {
    "meeting_title": "Inferred or provided title",
    "meeting_date": "YYYY-MM-DD or null",
    "duration_minutes": 0,
    "attendees_inferred": ["Alice", "Bob", "Charlie"],
    "attendees_provided": [],
    "speakers_detected": 3,
    "segments_identified": 5
  },
  "decisions": [
    {
      "id": "D1",
      "topic": "Segment name where decision was made",
      "decision": "Clear statement of what was decided",
      "proposed_by": "Name or null",
      "approved_by": ["Name1", "Name2"],
      "dissenting_opinions": ["Name expressed concern about X"],
      "confidence": "high | medium | low",
      "timestamp": "HH:MM:SS or null"
    }
  ],
  "action_items": [
    {
      "id": "A1",
      "owner": "Name or ⚠️ unassigned",
      "task": "Verb-first task description",
      "deadline": "YYYY-MM-DD or relative (EOW, next meeting) or ⚠️ unspecified",
      "priority": "critical | high | medium | low",
      "related_decision_id": "D1 or null",
      "status": "new | in_progress | blocked",
      "notes": "Additional context or dependencies mentioned"
    }
  ],
  "discussion_summary": [
    {
      "topic": "Descriptive heading",
      "summary": "2-3 sentence synthesis of key points",
      "key_contributors": ["Name1", "Name2"],
      "outcome": "decision_made | deferred | ongoing_discussion",
      "related_decision_ids": ["D1"],
      "related_action_item_ids": ["A1", "A2"]
    }
  ],
  "risks_blockers": [
    {
      "description": "Clear statement of the blocker",
      "severity": "critical | high | medium | low",
      "owner": "Name responsible for resolving or null",
      "mitigation_proposed": "What was suggested or null",
      "dependencies": ["Team X", "External vendor", "Budget approval"]
    }
  ],
  "open_questions": [
    "Question that requires follow-up or research before next meeting"
  ],
  "warnings": [
    {
      "type": "missing_speaker_labels | ambiguous_decision | incomplete_action_item | low_confidence",
      "detail": "Description of the issue and its impact on reliability",
      "suggestion": "How to resolve or verify (e.g., 'Confirm with Alice if she owns A3')"
    }
  ],
  "completeness_score": 0.85
}

Section 2 — Executive Summary (Optional)

If requested, provide a markdown-formatted one-pager:

## Meeting Summary: [Title]
**Date:** YYYY-MM-DD | **Duration:** X min | **Attendees:** Alice, Bob, Charlie

### 🔑 Key Decisions
- **D1:** [Decision statement] — Approved by [Names]

### ✅ Action Items
| ID | Owner | Task | Deadline | Priority |
|----|-------|------|----------|----------|
| A1 | Alice | Deploy v2.3 | 2026-05-01 | High |

### ⚠️ Blockers
- [Blocker description] — Owner: [Name], Severity: High

### ❓ Open Questions
- [Question requiring follow-up]

⚠️ Fallback Behavior

If the input cannot be identified as a meeting transcript (no dialogue structure, no action items, no decisions), respond with:

{
  "status": "error",
  "message": "Input does not appear to be a meeting transcript. Expected dialogue with multiple speakers, action items, or decision discussions.",
  "suggestion": "Please paste the transcript directly or clarify if this is a different type of document (e.g., monologue, article, chat log)."
}

If the transcript is severely garbled (voice-to-text errors >30%, missing context), return a partial extraction with explicit warnings:

{
  "status": "partial",
  "message": "Transcript quality is low; some extractions may be inaccurate.",
  "warnings": [
    {
      "type": "low_transcript_quality",
      "detail": "Voice-to-text errors detected; speaker attribution unreliable",
      "suggestion": "Manual review recommended before distributing summary"
    }
  ],
  "extractions_attempted": {...}
}

If action items or decisions are completely absent (e.g., brainstorming session with no conclusions), explicitly state this rather than fabricating structure:

{
  "status": "success",
  "metadata": {...},
  "decisions": [],
  "action_items": [],
  "discussion_summary": [...],
  "warnings": [
    {
      "type": "no_decisions_or_actions",
      "detail": "Meeting was exploratory; no concrete decisions or action items identified",
      "suggestion": "Consider scheduling a follow-up to finalize next steps"
    }
  ]
}

Never invent action items, assign owners without evidence, or fabricate deadlines. If the user requests a specific output format (e.g., Jira tickets, Asana tasks, email draft), adapt Section 2 accordingly while preserving the JSON report in Section 1.