Meeting Action Extractor

No connector required for core functionality — this skill performs pure LLM analysis of pasted transcripts. The Airtable connector above is optional and only activated if the user wants to push action items to a tracking base.

Purpose

Paste a meeting transcript from any source (Zoom, Google Meet, Otter.ai, manual notes) and receive structured output: decisions, action items with owners and deadlines, open questions, key insights, implicit commitments, and follow-ups needed.

Step 1: Ingest Transcript

Accept input in any format:

• Pasted text (raw or formatted)
• Meeting notes with timestamps
• Auto-generated transcripts (Zoom, Otter.ai, Google Meet)
• Informal notes or bullet points

If the transcript is too short or vague to extract meaningful content, say so clearly — do not fabricate items.

Step 2: Extract Intelligence

Parse the transcript and extract 6 categories:

1. Decisions Made

For each decision:

• What was decided
• Who made the call
• Context (what was discussed before the decision)
• Confidence: High (explicit statement) / Medium (implicit agreement) / Low (assumed from lack of objection)

2. Action Items

For each action:

• Task description (specific, actionable)
• Owner (who's responsible — extract from "I'll do X" or "John, can you X")
• Deadline (explicit date, or inferred: "by next week", "before the launch")
• Priority: P0 (blocking) / P1 (important) / P2 (nice to have)
• Dependencies (does this need something else to happen first?)

3. Open Questions

Unresolved items that need follow-up:

• The question
• Who raised it
• Who should answer
• Suggested deadline for resolution

4. Key Insights

Important information shared during the meeting:

• The insight
• Who shared it
• Relevance (why it matters)

5. Implicit Commitments

Things people said they'd do that weren't formally assigned:

• "I was thinking about doing X" → potential commitment
• "We should probably X" → unassigned action
• Flag these separately — they need explicit confirmation from the owner

6. Follow-Up Meetings Needed

• Topic
• Suggested attendees
• Suggested timeframe
• Why it's needed

Step 3: Confidence Scoring

For each extracted item, assign a confidence score:

• High (90%+): Explicit statement — "John will send the report by Friday"
• Medium (60-89%): Implied from context — discussed at length, seemed agreed
• Low (30-59%): Inferred — mentioned briefly, no explicit confirmation

Flag anything below 60% with ⚠️ for human verification.

Step 4: Output Format

## Meeting Summary: [Title/Date]
**Attendees:** [extracted names]
**Duration:** [if timestamps present]

### Decisions (X items)
| # | Decision | Made By | Confidence |
|---|----------|---------|------------|
| 1 | ... | ... | High ✅ |

### Action Items (X items)
| # | Task | Owner | Deadline | Priority | Status |
|---|------|-------|----------|----------|--------|
| 1 | ... | ... | Apr 10 | P1 | Pending |

### Open Questions (X items)
| # | Question | Raised By | Answer From | Deadline |
|---|----------|-----------|-------------|----------|

### Key Insights
- [insight 1 — who shared it]
- [insight 2]

### Implicit Commitments ⚠️
- [commitment — needs confirmation from owner]

### Follow-Up Meetings
- [Topic] — [attendees] — [timeframe]

Step 5: Push to Airtable (Optional)

If the user wants to track action items in Airtable:

use airtable AIRTABLE_CREATE_RECORDS '{"baseId":"appXXX","tableIdOrName":"Action Items","records":[{"fields":{"Task":"...","Owner":"...","Deadline":"...","Priority":"P1","Status":"Pending","Source Meeting":"[date]"}}]}'

Ask the user for their Airtable base ID and table name if not already configured.

Step 6: Follow-Up Email Draft

Generate a follow-up email summarizing:

• What was decided
• Who's doing what by when
• Open items that need resolution
• Next meeting (if any)

Tone: professional, concise, bullet-point format. Ready to send via Reed (Email Secretary).

Transcript Handling Tips

• Auto-generated transcripts have speaker labels — use them for owner assignment
• If no speaker labels are present, ask the user to clarify who said what for ambiguous ownership
• Handle multiple meetings in batch if provided
• Ignore small talk, pleasantries, and off-topic discussion — extract only actionable content
• Convert relative dates to absolute ("next Friday" → specific date based on meeting date)
• Group related action items under the decision they stem from where possible

Guidelines

• Never fabricate action items not present in the transcript
• If owner is unclear, mark as "TBD" and flag for human assignment
• If the transcript is too short or vague, say so — do not invent content
• Items below 60% confidence must be flagged with ⚠️ for human verification
• Batch processing is supported — handle multiple transcripts sequentially if provided