analyze-voice

name	analyze-voice
description	Analyze voice memo transcriptions, create notes with appropriate content, and return lightweight proposals. For resources (ideas, reflections), creates note with Layer 1 transcription. For meetings, extracts structured body content. Worker skill for triage orchestrator.
user-invocable	false

Analyze Voice Memo

Analyze a single voice memo transcription, create the appropriate note, and return a lightweight proposal.

Key design: This skill reads the full transcription AND creates the note before returning. For meetings, it extracts structured body content. For resources (ideas, reflections), it injects the transcription as Layer 1. The full content stays in subagent context - only the proposal flows back.

Input

You receive:

• file: Path to transcription in inbox (e.g., 00 Inbox/🎤 2024-01-22 3-45pm.md)
• areas: Available areas in vault
• projects: Available projects in vault
• stakeholders: Known people with names, roles, and companies (from config)

Output

Return a JSON proposal per @plugins/para-obsidian/skills/triage/references/proposal-schema.md.

Key: Use area (single wikilink), project (single wikilink or null), resourceType (camelCase). Include file, type: "transcription", created, and layer1_injected alongside the standard proposal fields.

For voice memos, always set source_format: "audio". For meetings, include meeting-specific fields (meeting_type, meeting_date, attendees, meeting_notes, decisions, action_items, follow_up).

Workflow

Step 1: Read Transcription

para_read({ file: "[input file]", response_format: "json" })

Extract frontmatter fields (recorded, summary, pre-filled areas/projects, and any existing metadata) from the YAML header in the para_read response. Do NOT call para_fm_get separately — para_read returns the full file including frontmatter.

Also extract the full transcription text from the body.

Step 2: Match Speakers to Stakeholders

CRITICAL: Match speaker names/aliases against the stakeholders list.

For each speaker mentioned in the transcription:

1. Check against stakeholder name (exact or partial match)
2. Check against stakeholder alias (e.g., "MJ" → "Mustafa Jalil")
3. Check against email prefix (e.g., "JXu3" → "June Xu")
4. If matched, output as wikilink: "[[June Xu]]"
5. If no match, output as plain text: "Speaker 3" or "Unknown (marketing team)"

Project inference: If ALL identified speakers share the same squad/project in stakeholders, infer that project:

Speakers: June Xu, Mustafa Jalil, Joshua Green
All in squad: "GMS (POS Yellow)"
→ suggested_projects: ["[[🎯 GMS - Gift Card Management System]]"]

Step 3: Analyze Content

Voice memos are inherently ambiguous. Analyze for:

1. Summary hint: If summary exists in frontmatter, use it as a quick signal for meeting type (e.g., "team planning session" → planning, "1:1 with manager" → 1on1)
2. Speaker roles: Matched stakeholders' roles help classify (e.g., conversation with "manager" → likely 1on1)
3. Speaker count: One person (reflection) vs multiple (meeting)?
4. Tone: Professional, personal, brainstorming?
5. Structure: Agenda items, action items, freeform?
6. Time markers: "Next week", "by Friday" suggest action items
7. Names mentioned: People, projects, tools?

Step 4: Categorize

Pattern	Resource Type	Meeting Type
Multiple speakers, status updates	meeting	standup
Two people, career/personal topics	meeting	1on1
Sprint planning, backlog discussion	meeting	planning
What went well/wrong, improvements	meeting	retro
Technical deep-dive, learning	meeting	workshop
Multiple speakers, general discussion	meeting	general
Single speaker, thinking aloud	idea	null
Single speaker, emotions/journaling	reflection	null
Discussion with back-and-forth	conversation	null

Step 5: Extract Body Content (for meetings)

Attendees: All speakers identified in Step 2 (wikilinks for matched, plain text otherwise).

Meeting Notes: Key discussion points, observations, notable comments. Look for:

• Status updates
• Problems discussed
• Ideas proposed
• Technical explanations

Decisions: Important decisions reached. Look for:

• "We decided..."
• "Let's go with..."
• "The plan is..."
• "Agreed that..."

Action Items: Extract structured tasks with:

• assignee: Wikilink if matched stakeholder, otherwise name or omit
• task: What needs to be done
• due: Date if mentioned (YYYY-MM-DD format), otherwise null

Look for:

• "I need to..."
• "We should..."
• "[Name] will..."
• "Action item:"
• "TODO:"
• "By [date]..."
• "Follow up on..."

Follow-up: Items for future discussion or preparation.

Step 5.5: Pre-Creation Self-Check

Before calling para_create, verify your args object includes these critical fields:

If meeting:

• summary — one-sentence description
• area — single wikilink
• meeting_type — standup, 1on1, planning, retro, workshop, general

If resource:

• summary — one-sentence description
• areas — wikilink(s)
• source_format — audio
• resource_type — idea, reflection, conversation

If any field is missing but you determined a value in Steps 3-5, add it to args now.

Step 6: Create Note (Single Call)

This is where content stays isolated. Use para_create with the content parameter for both meetings and resources:

Meeting-specific: Pass meeting body content (attendees, notes, decisions, action items, follow-up) via content parameter. Set layer1_injected: null for meetings.

para_create({
  template: "meeting",
  title: proposed_title,
  args: { ...meeting fields from validArgs },
  content: {
    "<attendees-section>": attendeesList,
    "<notes-section>": notesList,
    "<decisions-section>": decisionsList,
    "<action-items-section>": actionItemsList,
    "<follow-up-section>": followUpList
  },
  response_format: "json"
})

Resource-specific: Pass transcription as Layer 1 via content parameter. Set layer1_injected: true.

• If transcription <2k tokens: Include full transcription
• If transcription >2k tokens: Sample key segments using these strategies:
  • Include the opening (first 2-3 paragraphs for context)
  • Include sections with key insights (look for declarative statements, "the key thing is...", "what I learned...")
  • Include sections with action items or decisions
  • Include the closing (last 1-2 paragraphs for conclusion)
  • Preserve timestamps in sampled sections (e.g., [03:45])
  • Target 2-3k tokens total
• Always add: *Transcription captured. Use /distill-resource to extract key insights.*

para_create({
  template: "resource",
  title: proposed_title,
  args: { ...resource fields from validArgs },
  content: {
    "<content-target-heading>": formattedTranscription
  },
  response_format: "json"
})

No separate para_commit or para_replace_section calls needed — the CLI handles injection and commit internally.

IMPORTANT: Do NOT archive or delete the original transcription. Cleanup is the coordinator's responsibility (Phase 5, after user review).

Step 7: Verify & Repair

During triage: Skip this step entirely. Set verification_status: "pending_coordinator" and verification_issues: []. The coordinator handles verification in Phase 2.5 — it stamps and checks all critical fields from PROPOSAL_JSON.

Standalone callers (non-triage): Call para_fm_get on the created file. Check the critical fields for your template (meeting or resource) using the same table from Step 5.5. If any MISMATCH: para_fm_set({ file, set: { ...repairs }, response_format: "json" }). Set verification_status: "verified", "repaired", "needs_review", or "skipped".

Step 8: Return Proposal

Return the lightweight JSON proposal. The note is already created.

Voice Memo Challenges

Voice memos are the hardest to categorize because:

1. No clear boundaries - Conversations meander
2. Context missing - You don't know the full situation
3. Multiple topics - One recording may cover several things
4. Transcription errors - Names, technical terms may be garbled

When in doubt, set confidence: "low" so the user can choose "Deeper" for multiple options.

Confidence Levels

Level	When to Use
high	Clear meeting format, obvious participants, structured content
medium	Likely interpretation but could be wrong
low	Ambiguous content, multiple valid interpretations

Default to medium or low for voice memos. They're inherently ambiguous.

Example Output

Meeting Proposal (note already created)

{
  "file": "00 Inbox/🎤 2024-01-22 3-45pm.md",
  "type": "transcription",
  "proposed_title": "Sprint 42 Planning Session",
  "proposed_template": "meeting",
  "summary": "Team planning session discussing priorities for Sprint 42. Focus on completing the migration project and addressing tech debt in the auth module.",
  "created": "04 Archives/Meetings/Sprint 42 Planning Session.md",
  "layer1_injected": null,
  "verification_status": "verified",
  "verification_issues": [],
  "categorization_hints": [
    "Migration project is highest priority",
    "Auth module tech debt blocking other work",
    "New hire starting next week needs onboarding"
  ],
  "area": "[[🌱 Work]]"
   
        
         
       
   
   
   

     
   
    
    
    
    
  
   
    
    
    
  
   
           
           
           
  
   
    
    
  

Resource Proposal (voice memo → idea, note already created with Layer 1)

{
  "file": "00 Inbox/🎤 2024-01-20 8-30am.md",
  "type": "transcription",
  "proposed_title": "API Gateway Caching Idea",
  "proposed_template": "resource",
  "summary": "Quick brainstorm about implementing response caching at the API gateway level to reduce database load.",
  "created": "03 Resources/API Gateway Caching Idea.md",
  "layer1_injected": true,
  "categorization_hints": [
    "Redis-based caching at gateway",
    "TTL based on endpoint patterns",
    "Could reduce DB load by 40%"
  ],
  "area": "[[🌱 Work]]",
  "project": "[[🎯 Performance Optimization]]",
  "resourceType": "idea",
  "source_format" 
   
   
   
   
   
   

Deep Analysis Mode

When the coordinator requests "deeper" analysis, return 3 different interpretations:

{
  "options": [
    {
      "label": "A",
      "interpretation": "Sprint Planning Meeting",
      "proposed_title": "Sprint 42 Planning",
      "proposed_template": "meeting",
      "resourceType": "meeting",
      "meeting_type": "planning",
      "rationale": "Multiple speakers, task assignments, timeline discussion",
      "attendees": ["[[June Xu]]", "[[Mustafa Jalil]]"],
      "meeting_notes": ["..."],
      "decisions": ["..."],
      "action_items": [       
    
    
       
       
       
       
       
       
       
    
    
       
       
       
       
       
       
       
        
        
       
       
    
  

This gives the user meaningful choices when categorization is ambiguous. Each option includes full body content extraction where applicable.