Meeting Notes to Action Items

What you'll build

A pipeline that converts meeting notes or a transcript into:

• Decisions — what the group actually agreed
• Action items — owner, task, due date, priority
• Open questions — unresolved threads with suggested owners
• Risks / blockers — what could derail the work

Then optionally syncs actions to Linear/Jira and posts a Slack recap. The goal isn't a prettier summary — it's fewer dropped commitments.

The gap transcription tools don't close

Otter, Meet, and Zoom give you text. They might even bullet "key takeaways." They rarely:

• Assign owners correctly when names are implicit ("Sarah's team will handle it")
• Distinguish decisions from brainstorm ideas
• Produce tracker-ready tasks with due dates
• Post to Slack with @mentions

LLM extraction with a locked JSON schema plus a 2-minute human check closes that gap for ~$0.01/meeting.

Architecture

Transcript / notes
  → Clean (strip small talk)
    → LLM extraction (JSON)
      → Ambiguity flags (optional)
        → Human edit
          → Markdown recap + Linear API + Slack

Native shortcuts:

• Gemini in Google Meet for live notes + Workspace recap
• ChatGPT for paste-and-run with the meeting-notes prompt
• Custom CLI when you want Linear sync and audit logs

Step 1 — Input

Minimum viable: bullet notes you typed during the call.

Better: auto-transcript with speaker labels.

Pre-process:

• Remove AV check, small talk first 2 minutes
• Normalize names (Jon → Jonathan Park via attendee list)
• If transcript is > 20K words, split by agenda section

Step 2 — Schema

Use structured output everywhere — same shape, every meeting:

{
  "decisions": [{ "text": "...", "source_quote": "..." }],
  "actions": [{
    "owner": "Alex Chen",
    "task": "Send pricing draft to legal",
    "due": "2026-06-10",
    "priority": "high",
    "source_quote": "Alex will get legal the pricing draft by Tuesday"
  }],
  "open_questions": [{ "text": "...", "suggested_owner": "unassigned" }],
  "risks": [{ "text": "...", "severity": "medium" }]
}

source_quote is your hallucination alarm — if the quote doesn't exist in the notes, delete the row.

Step 3 — Extract

Run the structured extraction prompt (frontmatter) with:

• Full notes/transcript
• Attendee list (critical for owners)
• Optional meeting type (standup | planning | retro) to tune emphasis

Standup: favor blockers and today's commits.
Planning: favor decisions and dated milestones.
Retro: favor risks and process actions.

Model pick: Gemini Flash or Haiku for daily; Sonnet when the meeting was argumentative or ambiguous.

Step 4 — Validate

Non-negotiable 2-minute checklist:

• Every action has the right owner (not "the team")
• Due dates are real or explicitly TBD
• Decisions aren't duplicated as actions
• Vague tasks rewritten ("look into API" → "Spike: document rate limits for Partner API")

Run the ambiguity pass prompt when >30% of owners are unassigned.

Step 5 — Sync

Slack recap — format JSON with the Slack prompt; @mention owners.

Linear / Jira — one issue per action:

for (const action of extraction.actions) {
  if (action.owner === "unassigned") continue; // or route to triage
  await linear.createIssue({
    teamId: TEAM,
    title: action.task,
    assigneeId: resolveUser(action.owner),
    dueDate: action.due === "TBD" ? undefined : action.due,
    description: `From meeting ${meetingId}\n\n> ${action.source_quote}`,
  });
}

Doc archive — save transcript + JSON to meetings/2026-06-06-product-sync.md for search.

Meeting-type templates

Daily standup

Extract only:

• Yesterday done / today doing / blockers
• Skip decisions unless explicit

Planning / review

Extract:

• Decisions with alternatives considered
• Dated actions tied to milestones
• Open questions blocking estimates

Retro

Extract:

• Process changes (actions)
• Risks (systemic, not blame)
• Decisions to try an experiment next sprint

Privacy and compliance

• Redact customer names and dollar figures before LLM if policy requires
• Use enterprise endpoints (Gemini Vertex, ChatGPT Team, Claude Team) for HR/legal
• Don't store raw audio in unsecured drives — transcript text only
• Retention policy: delete transcripts after 90 days if compliance demands

Extending

• Pre-meeting brief — pull open actions from Linear, attach to agenda
• Follow-up bot — Friday Slack DM listing open actions assigned to you
• Cross-meeting dedup — hash task text, merge "send legal the draft" across two meetings
• Email mode — same pipeline for client call notes → AI Email Assistant draft

Operating cost

10 meetings/week × $0.01 ≈ $0.40/month in inference. Even with Otter at $20/month, total cost is trivial vs one forgotten action item delaying a launch.

Pairs well with

• Gemini + Google Meet for capture
• NotebookLM for turning a quarter of meeting notes into study/reflection material — different job
• best AI productivity tools for picker context
• Linear/Jira as system of record — extraction without sync is where workflows die