process-meeting

name	process-meeting
effort	medium
description	Process Google Meet transcripts from Google Drive OR Apple Voice Memos into structured Obsidian vault notes with topic extraction, vault linking, Todoist action items, and calendar deadlines. Invoke with `/process-meeting`, `/process-meeting latest`, `/process-meeting YYYY-MM-DD`, or `/process-meeting voice [latest|today|YYYY-MM-DD]` for Apple Voice Memo recordings transcribed locally via mlx-whisper. Also trigger when user says "process meetings", "pull in transcripts", "process today's meeting", "process voice memo", "transcribe my memo", or asks to import meeting notes from Google Drive or voice memos.

Process Meeting

Transform Google Meet recordings (Gemini notes with embedded transcripts) OR Apple Voice Memo recordings (transcribed locally via mlx-whisper) into structured Obsidian vault notes. Extracts topics with what/why/context narratives, links to existing vault knowledge, creates Todoist action items, and adds calendar deadlines.

Source modes

Two pipelines, picked by subcommand:

• Default (Google Meet): pulls Gemini notes from Google Drive via rclone. See Google Meet Pipeline below.
• voice subcommand (Apple Voice Memos): reads .m4a files from the iCloud-synced Voice Memos folder, transcribes with mlx-whisper, then routes to a solo or multi-person template. See Voice Memo Pipeline below.

Google Meet Pipeline

Run stages sequentially. Each stage is a checkpoint — if a later stage fails, earlier work is preserved.

Fetch → Parse → Write → Link → Action → Calendar

Time awareness: Before starting, check the current date and time (run date). Sessions can span date boundaries — do not rely on the date provided at session start. Use the freshly checked date for all references to "today", relative dates, and temporal language in output.

Stage 1: Fetch

Configured sources: Read from config.json in this skill's directory. Each source has a name, remote (rclone path), and flags array. See config.example.json for the schema.

Iterate over all sources from config. First list files with rclone lsf, then copy matches as plain text.

CRITICAL — Date format in filenames: Google Drive filenames use fullwidth slashes (／, Unicode U+FF0F) as date separators, NOT regular slashes or hyphens. A file dated April 2, 2026 appears as 2026／04／02 in the filename. You MUST use fullwidth slashes in all --include patterns and grep filters. Standard hyphens (-) or regular slashes (/) will silently match nothing.

To list files for a given date:

rclone lsf "<remote>" [flags] | grep "2026／04／02"

To fetch files for a given date:

rclone copy --drive-export-formats txt "<remote>" /tmp/meet-staging/ --include "*2026／04／02*" [flags]

Filter files to only those containing "Notes by" (Gemini meeting notes with embedded transcripts).

Argument handling:

• No argument: default to today (current date only)
• all: process all unprocessed meetings from all sources (warn user about volume first)
• latest: most recent file only (across all sources)
• today: resolve to current date, then use fullwidth slash format YYYY／MM／DD for matching
• YYYY-MM-DD: convert to fullwidth slash format YYYY／MM／DD for matching
• week: last 7 days
• month: last 30 days

When no argument is provided and no meetings are found for today, ask the user: "No meetings found for today. How far back should I look?" and offer options: latest, week, month, or a specific date.

Deduplication: Scan 2_Literature Notes/ for existing meeting notes. Match by date AND participant names to avoid collisions (multiple meetings on the same day across sources). Skip any meeting that already has a corresponding note.

Important: Fullwidth slashes (／) in Google Drive filenames break rclone cat. Always use rclone copy with --include glob patterns, never rclone cat with fullwidth-slash filenames.

Stage 2: Parse

Use the full transcript (everything after the Transcript marker), NOT Gemini's summary. The transcript is the source of truth — it captures the reasoning and context that Gemini loses.

For each meeting:

1. Identify distinct topics from conversation flow and timestamp markers (00:09:56 format). A new topic = substantive subject shift.

2. For each topic extract:

   • Title (concise, descriptive)
   • Timestamp range
   • Narrative summary: what was discussed, why it matters, situational context — as a coherent paragraph, not a form. Capture reasoning, stories, analogies, and evidence behind decisions.
   • Decisions made (bulleted)
   • Open questions (bulleted)

3. Extract action items with owners and due dates if mentioned

4. Extract hard deadlines — conferences, launches, external commitments with dates

Participant names: Parse from transcript speaker labels (e.g., Jonah Harris:, Joe Cotellese:).

Stage 3: Write

Create note at: 2_Literature Notes/YYYYMMDD - <Participants> Meeting - <Brief Description>.md

See references/note-template.md for the complete template.

Key rules:

• Brief description: derive from 2-3 most significant topics, under 60 characters
• Domain tags: infer from content. Common: nextgres, business, engineering, productmanagement, ai, database
• Participants as [[wiki-links]] in frontmatter

Stage 4: Link

After writing the note:

1. Search vault (3_Permanent Notes/, 2_Literature Notes/, 4_Projects/) for concepts mentioned in the meeting
2. Add [[wiki-links]] inline where matches exist
3. Add "Suggested Permanent Notes" section at the end — topics that might deserve their own evergreen note in 3_Permanent Notes/. One-line rationale each. Do NOT auto-create.

Stage 5: Action Items → Routing (Notion / Todoist)

Only consider tasks owned by the vault owner (Joe Cotellese). Other participants' action items should appear in the meeting note but NOT be pushed anywhere.

Routing rule — classify each task before creating it:

• NEXTGRES-related → NEXTGRES Notion Kanban Board (see below). A task is NEXTGRES-related if it concerns the NEXTGRES product/company (customers, investors, the Postgres proxy, its roadmap, etc.), or the source note carries the nextgres domain tag and the task isn't clearly personal.
• Everything else → Todoist inbox.

When a task is ambiguous, default to Todoist.

1. Present proposed tasks using AskUserQuestion with multiSelect: true. Each option:

   • label: The task title (concise, actionable)
   • description: Standalone context + the routing destination in brackets, e.g. [→ Notion] or [→ Todoist], so the user can veto a misroute.

2. Create only selected tasks. If the user selects none, skip creation entirely.

3. If no Joe action items were extracted, skip this stage entirely — do not prompt.

Common fields for every task:

• title/content: Clear, actionable task title
• description/body: Standalone context + Obsidian link: obsidian://open?vault=obsidian-vault&file=2_Literature%20Notes%2F<filename>.md
• due date: From transcript context if a date/timeframe is mentioned, otherwise omit

NEXTGRES tasks → Notion. Read notionRouting from config.json (this skill's directory; see config.example.json for the schema). If that block is missing, warn and fall back to Todoist for these tasks — do not hardcode IDs here. Create via notion-create-pages with parent: {type: "data_source_id", data_source_id: "<notionRouting.dataSourceId>"} and properties:

• Task name: the title
• Assignee: ["<notionRouting.assigneeUserId>"] ("tagged with me")
• Status: <notionRouting.defaultStatus>
• date:Due date:start: ISO date if one was mentioned, otherwise omit
• content: the standalone context + Obsidian link (page body)

Other tasks → Todoist via Todoist MCP add-tasks, destination inbox (no projectId), with dueString when a timeframe was mentioned.

Stage 6: Calendar (iMCP)

Only for hard deadlines — conferences, launches, contractual dates.

• Create calendar event via iMCP MCP tools
• Add notice period scaled to urgency (1 week for month-out, 2-3 days for week-out)
• If iMCP not connected: Warn user with list of deadlines needing manual entry. Do NOT fail.

Completion Output

## Meeting Processing Complete

**Processed:** <count> meeting(s)
**Notes created:**
- [Title](obsidian://open?vault=...) — <topic count> topics

**Todoist tasks created:** <count>
**Calendar events:** <count> (or "iMCP offline — <count> deadlines need manual entry")

**Suggested Permanent Notes:**
- <topic> — <rationale>

---

Voice Memo Pipeline

Triggered by /process-meeting voice [latest|today|YYYY-MM-DD]. Two paths inside this pipeline: solo (notes-to-self) and multi-person (recorded conversation). See references/voice-memo-template.md for the templates and classification heuristic.

Fetch → Transcribe → Classify → Parse → Write → Link → Action → Calendar

Stage 1: Fetch

Source folder: ~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/

Filename format: YYYYMMDD HHMMSS[-HEXID].m4a (Apple-assigned). User-renamed memos may use arbitrary names — accept both. Date is parsed from the leading YYYYMMDD token when present; otherwise fall back to the file's mtime.

List recordings:

ls "$HOME/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/"*.m4a

Argument handling:

• voice (no further arg): default to today
• voice latest: most recent .m4a by mtime
• voice today: today's date in YYYYMMDD prefix match
• voice YYYY-MM-DD: specific date — convert to YYYYMMDD prefix match
• voice all: all unprocessed (warn about volume first)

Deduplication: Scan 1_inbox/ (solo path) and 2_Literature Notes/ (multi-person path) for any note whose frontmatter source_file matches the .m4a basename. Skip already-processed.

Stage 2: Transcribe

Run mlx-whisper via the helper script:

~/.claude/skills/process-meeting/scripts/transcribe-voice.sh \
  "<input.m4a>" \
  /tmp/voice-staging/

The script outputs /tmp/voice-staging/<basename>.txt. Default model is mlx-community/whisper-large-v3-mlx (override via MLX_WHISPER_MODEL env var). The script exits with code 69 if mlx_whisper is not on PATH — install with uv tool install mlx-whisper.

Long recordings: Run transcription in the background (run_in_background: true) for files >5 minutes; the model takes roughly real-time on Apple Silicon. Don't poll — wait for completion notification.

Stage 3: Classify

Apply the heuristic in references/voice-memo-template.md:

• Single first-person narrator throughout → solo path
• Question/answer turns or proper-name addressing ("Hey Jonah") → multi-person path
• Ambiguous / short (<60s) → default solo, but ask if uncertain

When uncertain, ask the user with AskUserQuestion showing a 2-3 sentence transcript excerpt for context.

Stage 4: Parse

Solo path:

• Identify the through-line (the dominant subject the user is thinking about)
• Extract: 2-4 sentence summary, key thoughts (bullets), decisions, action items, open questions
• Title: under 60 characters, derived from the dominant subject

Multi-person path:

• Same as Google Meet Stage 2 (topics + narrative + decisions + open questions + action items)
• Speaker attribution: mlx-whisper does NOT diarize. Infer participants from name mentions ("hey Jonah", "Joe asked"). If unclear, prompt user.
• Replace timestamp lines with *Approx: <ordinal>* (e.g., "Approx: opening third") or omit.

Stage 5: Write

Solo path: 1_inbox/YYYYMMDDHHMMSS - Voice Memo - <Title>.md — see solo template in references/voice-memo-template.md. Tag includes voice-memo.

Multi-person path: 2_Literature Notes/YYYYMMDD - <Participants> Meeting - <Brief Description>.md — same as Google Meet path but with source: "Apple Voice Memo - mlx-whisper transcript" and added source_file/duration frontmatter fields.

Stage 6: Link

Same as Google Meet Stage 4 — search vault, add [[wiki-links]], append "Suggested Permanent Notes" section. Solo memos may yield fewer links because they're more fleeting; that's fine.

Stage 7: Action Items → Routing (Notion / Todoist)

Same Joe-only rule and NEXTGRES-vs-Todoist routing as Google Meet Stage 5. Solo memos often surface latent todos ("I should call X") — capture those, routing NEXTGRES items to the Notion Kanban board and everything else to the Todoist inbox, each with the Obsidian deep link in the description.

Stage 8: Calendar (iMCP)

Same as Google Meet Stage 6. Voice memos rarely contain hard deadlines, but if one is dictated, treat identically.

Voice Memo Completion Output

## Voice Memo Processing Complete

**Processed:** <count> recording(s)
**Path:** <solo|multi-person|mixed>
**Notes created:**
- [Title](obsidian://open?vault=...) — <path>

**Todoist tasks created:** <count>
**Calendar events:** <count>