Extract Action Items

Extract action items from a Fireflies transcript using parallel subagents. Catches items automated summaries miss.

Two modes:

• All attendees (default): No target specified — extract action items for every participant
• Single person: Target specified — extract action items for that person only

Phase 1: Determine Mode

Parse the user's invocation:

• If a target person is specified → single-person mode
• Otherwise → all-attendees mode

Extract the search criteria (date, keyword, or transcript ID) from the invocation.

Phase 2: Fetch & Preprocess (Subagent)

The transcript API returns a JSON array (or an MCP wrapper containing one). Extract to plain text before chunking.

You should inspect the user's local hooks config and avoid running commands that are blocked by the hooks.

MCP based extraction

mkdir -p .claude/scratchpad
node -e "
  const fs = require('fs');
  let data = JSON.parse(fs.readFileSync(process.argv[1], 'utf8'));
  // Handle MCP wrapper: if top-level array has a .text field containing the real transcript, parse that
  if (data.length === 1 && typeof data[0]?.text === 'string') {
    // Extract speaker lines from the text content
    const lines = data[0].text.split('\n').filter(l => l.match(/^[A-Za-z].*?:/));
    fs.writeFileSync('.claude/scratchpad/transcript.txt', lines.join('\n'));
    const speakers = [...new Set(lines.map(l => l.split(':')[0].trim()))].sort();
    console.log('Speakers:', JSON.stringify(speakers));
    console.log('Total lines:', lines.length);
  } else {
    // Standard array of {speaker_name, text} objects
    const lines = data.map(e => (e.speaker_name || 'Unknown') + ': ' + (e.text || ''));
    fs.writeFileSync('.claude/scratchpad/transcript.txt', lines.join('\n'));
    const speakers = [...new Set(data.map(e => e.speaker_name).filter(Boolean))].sort();
    console.log('Speakers:', JSON.stringify(speakers));
    console.log('Total lines:', lines.length);
  }
" [TRANSCRIPT_JSON_FILE]

If the transcript JSON was saved to a tool-results file by the MCP client, pass that file path as the argument.

API based extraction

CRITICAL: The orchestrator MUST NOT call any Fireflies MCP tools directly. ALL Fireflies interaction happens inside this subagent.

Launch a single general-purpose subagent with this prompt:

Search Fireflies for a transcript matching: [SEARCH_CRITERIA]

1. Call `mcp__fireflies__fireflies_get_transcripts` to find the transcript (by date, keyword, or ID).
2. Call `mcp__fireflies__fireflies_get_summary` and `mcp__fireflies__fireflies_get_transcript` in parallel for the matched transcript.
3. The transcript API returns a JSON array. Extract to plain text:
   - With jq: jq -r '.[].text' < raw_transcript.json > .claude/scratchpad/transcript.txt
   - Fallback: python3 -c "import json,sys; print('\n'.join(e['text'] for e in json.load(sys.stdin)))" < raw_transcript.json > .claude/scratchpad/transcript.txt
4. Count lines: wc -l < .claude/scratchpad/transcript.txt
5. Extract the distinct speaker list from the transcript JSON:
   python3 -c "import json,sys; data=json.load(sys.stdin); print('\n'.join(sorted(set(e.get('speaker_name','') for e in data if e.get('speaker_name')))))" < raw_transcript.json

Return EXACTLY this (no other text):
- meeting_title: <title>
- meeting_date: <date>
- transcript_id: <id>
- transcript_path: .claude/scratchpad/transcript.txt
- line_count: <number>
- speakers: <comma-separated list>
- summary: <the Fireflies summary text>

Wait for the subagent to finish. Parse its returned values — these are the inputs for the remaining phases.

Phase 3: Parallel Subagent Extraction

Chunk sizing: ceil(total_lines / 5) lines per chunk, minimum 200. Adjust chunk count so no chunk is under 200 lines.

Launch one general-purpose subagent per chunk.

Single-Person Prompt

Read lines [START] to [END] of [FILE_PATH].

Find ALL action items for [TARGET_PERSON]. Return each as:
- **Item**: what they committed to
- **Quote**: exact words from transcript
- **Context**: who else involved, any deadline
- **Discussion depth**: If this item emerged from extended back-and-forth (design decisions, technical debates, multi-speaker deliberation), include: what was proposed, what alternatives were considered, what was decided and WHY, specific technical details (field names, schema choices, API behaviors), open questions or deferred items, and connections to other people's work

Beyond obvious commitments ("I'll do X"), catch these non-obvious patterns:
- Self-notes: "I'll make a note to...", "let me jot down..."
- Admissions implying catch-up: "I dropped the ball on X", "I still haven't read X"
- Conditional offers that became commitments: "If we have time, I'm happy to..."
- Volunteering: "I guess I'll volunteer to..."
- Exploration tasks: "Let me spend a few hours with it"
- Questions/topics for external parties: "I need to ask [person/firm] about X", "thing to discuss with [party]"

All-Attendees Prompt

Read lines [START] to [END] of [FILE_PATH].

The meeting attendees are: [SPEAKER_LIST]

Find ALL action items for EVERY attendee. Group by person. For each item return:
- **Person**: who owns the action item
- **Item**: what they committed to
- **Quote**: exact words from transcript
- **Context**: who else involved, any deadline
- **Discussion depth**: If this item emerged from extended back-and-forth (design decisions, technical debates, multi-speaker deliberation), include: what was proposed, what alternatives were considered, what was decided and WHY, specific technical details (field names, schema choices, API behaviors), open questions or deferred items, and connections to other people's work

Beyond obvious commitments ("I'll do X"), catch these non-obvious patterns:
- Self-notes: "I'll make a note to...", "let me jot down..."
- Admissions implying catch-up: "I dropped the ball on X", "I still haven't read X"
- Conditional offers that became commitments: "If we have time, I'm happy to..."
- Volunteering: "I guess I'll volunteer to..."
- Exploration tasks: "Let me spend a few hours with it"
- Questions/topics for external parties: "I need to ask [person/firm] about X", "thing to discuss with [party]"
- Delegations: "[Person], can you handle X?", "I'll leave that to [person]"

Phase 4: Synthesize Notes

Merge subagent results, deduplicate, and categorize into a rich synthesized notes file. This is the master working document — all detail lives here. Linear proposals and the final action items checklist are derived from it.

Write to .claude/scratchpad/synthesized-notes-YYYY-MM-DD.md. Only include categories that have items.

Synthesis Depth

Preserve the full Discussion depth returned by subagents. Never flatten discussion-rich items into one-liners.

• Checkbox title = the deliverable. Body = full context needed to execute it.
• If a subagent returned multi-paragraph context for an item, keep it. Use bold sub-headers to organize (e.g., "Root cause:", "Agreed approach:", "Open items:").
• Never collapse N distinct decisions into 1 bullet. List each.
• Cross-link items that depend on each other (e.g., "dependency for Emerson's fiscal period table work").
• Simple items (credential sharing, quick investigations) stay as one-liners.
• Include exact quotes from the transcript for each item.

Categories

1. High Priority / Technical — Code changes, bug fixes, PR reviews, investigations
2. Pairing / Collaboration — Scheduled syncs, joint work sessions
3. Content / Research — Reading, writing, experiments, documentation
4. Questions for External Parties — Topics to raise with specific people/firms outside the immediate team
5. Exploration / Tooling — Tool evaluations, setup, environment tasks
6. Catch-up — Things explicitly acknowledged as dropped or missed

Output Format

Single-person mode:

# [Name] Synthesized Notes — [Meeting Title]

**Date:** [Date]
**Fireflies Link:** https://app.fireflies.ai/view/[TRANSCRIPT_ID]

## [Category Name]

- **Item title**
  - Context, decisions, and full detail
  - > "Exact quote"

All-attendees mode:

# Synthesized Notes — [Meeting Title]

**Date:** [Date]
**Fireflies Link:** https://app.fireflies.ai/view/[TRANSCRIPT_ID]

## [Person Name]

### [Category Name]

- **Item title**
  - Context, decisions, and full detail
  - > "Exact quote"

Phase 5: Linear Ticket Proposals

Derive Linear ticket creates and updates from the synthesized notes. The rich context and quotes from Phase 4 flow into Linear (as comments or ticket descriptions) so it becomes the source of truth. Uses a config file for team defaults and queries active cycle tickets for update candidates.

5a: Config Resolution

Look for team configuration in this order (first match wins):

1. ~/.agents/configs/extract-my-action-items/config.json (user overrides)
2. references/config.json (bundled defaults, relative to this skill file)

Use the user config if found. Otherwise fall back to the bundled config.json.

If no user config exists AND the bundled config has an empty team field, stop and prompt the user:

No Linear config found. Create a user config at: ~/.agents/configs/extract-my-action-items/config.json

Copy the bundled references/config.json as a starting point and fill in your team, project, assignee, and labels.

If config resolves successfully, proceed.

5b–5c: Pull Active Tickets and Semantic Match (Single Subagent)

CRITICAL: Run 5b and 5c together inside a single general-purpose subagent. The cycle ticket data is large and should NOT flow through the main context window.

Launch a subagent with this prompt:

## Task: Pull active Linear tickets and match against synthesized meeting notes

### Step 1: Pull active tickets

Config: team=[TEAM], states=[STATES_LIST], attendees=[SPEAKER_LIST]

1. `mcp__linear__list_teams` with query=[TEAM] → get team ID
2. `mcp__linear__list_cycles` with type="current" → get current cycle ID
3. In parallel:
   - `mcp__linear__list_issues` filtered by cycle + team (limit 250)
   - `mcp__linear__list_issues` for each attendee (assignee filter, state="In Progress")
4. Deduplicate and build a lookup table: {identifier, title, assignee, status}

### Step 2: Semantic matching

Read the synthesized notes at [SYNTHESIZED_NOTES_PATH].

For each item, classify as:
- **UPDATE [TICKET-ID]** — maps to an existing ticket. Explain what new info to append.
- **NEW TICKET** — distinct deliverable not covered. Suggest title, assignee, priority.
- **IDEA** — process improvement, behavioral commitment, or exploratory thought.

Group output by classification. For UPDATE items include ticket ID. For NEW TICKET items include suggested title, assignee, and priority.

5d: Draft Proposals to Scratchpad

Write to .claude/scratchpad/linear-proposals-YYYY-MM-DD.md using the template from references/ticket-template.md.

• Proposed Updates: For each UPDATE match, draft a comment body with the new feedback (dated section with context and quotes from the synthesized notes). Do NOT modify the issue description — updates are posted as comments.
• Proposed New Tickets: Use send-to-linear description format (User Story, Requirements, Acceptance Criteria) with concrete examples and exact quotes from the synthesized notes.
• Ideas / Needs More Thought: List with person, context, and exact quote. These are not skipped — they appear in the proposals file but do not become full tickets.

5e: User Review Gate

STOP. Tell the user the proposals file is ready at .claude/scratchpad/linear-proposals-YYYY-MM-DD.md and wait for explicit instruction.

Use AskUserQuestion: "Linear ticket proposals are ready. Review the file, then choose:"

• "Create/update tickets in Linear" — proceed to execute
• "Skip — just do Slack DMs" — skip to Phase 7

The user may edit the scratchpad file before approving. On approval:

1. Resolve team ID, label IDs, project ID, and current cycle via Linear MCP (same pattern as send-to-linear Phase 6):
   • mcp__linear__list_teams → team ID
   • mcp__linear__list_issue_labels → label IDs
   • mcp__linear__list_projects → project ID (if configured)
   • mcp__linear__list_cycles with type: "current" → current cycle
2. For updates: mcp__linear__create_comment with issueId and the drafted comment body. Do NOT use mcp__linear__save_issue to modify the description.
3. For new tickets: mcp__linear__save_issue with all fields from config + proposal (team, project, assignee, cycle, state, labels, title, description)
4. Ideas — no Linear action (they stay in the proposals file for reference only)
5. Report results with clickable links so the user can verify:
   • Updated tickets: https://linear.app/[WORKSPACE]/issue/[TICKET-ID] for each commented ticket
   • Created tickets: https://linear.app/[WORKSPACE]/issue/[TICKET-ID] for each new ticket (use the identifier returned by save_issue)
   • Derive [WORKSPACE] from the team's organization key, or from the config if available

Phase 6: Action Items Checklist

Generate a terse action items checklist derived from the synthesized notes. Linear is the source of truth for detail — the checklist is just a scannable index with links.

Where an item maps to a Linear ticket (updated or created in Phase 5), include the Linear link inline. Items not sent to Linear get a one-line description only.

Output

Single-person mode — Write to .claude/scratchpad/[name]-action-items-YYYY-MM-DD.md:

# [Name] Action Items — [Meeting Title]

**Date:** [Date]
**Fireflies Link:** https://app.fireflies.ai/view/[TRANSCRIPT_ID]

## [Category Name]

- [ ] **Item title** — [TICKET-ID](https://linear.app/[WORKSPACE]/issue/[TICKET-ID])
- [ ] **Item without ticket** — brief context

All-attendees mode — Write to .claude/scratchpad/action-items-YYYY-MM-DD.md:

# Action Items — [Meeting Title]

**Date:** [Date]
**Fireflies Link:** https://app.fireflies.ai/view/[TRANSCRIPT_ID]

## [Person Name]

### [Category Name]

- [ ] **Item title** — [TICKET-ID](https://linear.app/[WORKSPACE]/issue/[TICKET-ID])
- [ ] **Item without ticket** — brief context

## Quick Reference — Time-Sensitive

1. [Person] — [Item with deadline]

Keep each item to one line.

Phase 7: Review & DM to Slack

1. Use AskUserQuestion: "DM action items to each person on Slack?" — options: "Send DMs", "Skip — just keep the file"
2. If approved, ensure .claude/slack-users.local.json exists in the project root:
   • If missing, run node [SKILL_DIR]/scripts/fetch-slack-users.mjs (requires SLACK_BOT_TOKEN with users:read scope), present the output to the user for review, then save to .claude/slack-users.local.json (gitignored by **/.claude/**/*.local.json)
   • If present, proceed directly
3. Run the bundled script with the output file path:

node [SKILL_DIR]/scripts/slack-post.mjs [OUTPUT_FILE_PATH]

The script sends Block Kit–formatted DMs to each person via conversations.open + chat.postMessage. Requires env var SLACK_BOT_TOKEN (with chat:write and im:write scopes).

Behavior by mode:

• All-attendees: Each person matched in slack-users.local.json receives a DM with only their action items. Unresolvable names are skipped with a warning.
• Single-person: One DM to the target person.

Name resolution supports exact match and fuzzy first-name match (e.g., "Jelvin" resolves to "Jelvin Base"). After the script runs, report any skipped names to the user.

4. After posting (or skipping), delete all artifacts created during the run: transcript.txt, synthesized-notes-YYYY-MM-DD.md, the action items markdown file, linear-proposals-YYYY-MM-DD.md, and any other temp files written to .claude/scratchpad/ during this workflow.

Example Invocations

• /extract-my-action-items — all attendees, most recent meeting
• /extract-my-action-items standup — all attendees, search for "standup"
• /extract-my-action-items for Basti from yesterday — single person
• /extract-my-action-items 01KFY1RSEVVQW7MB1TKG4N2D20 — all attendees, specific transcript