Minutes Record

Stop recording:

minutes stop

This sends a signal to the recording process, which then:

1. Stops audio capture
2. Transcribes the audio locally via whisper.cpp or parakeet.cpp (no cloud, no data leaves the machine)
3. Saves the transcript as a markdown file in ~/meetings/
4. Prints the output path and word count as JSON

Live transcript during recording:

While recording, Minutes streams a real-time transcript to ~/.minutes/live-transcript.jsonl. You can read it with:

minutes transcript                    # all lines
minutes transcript --since 42         # lines after cursor
minutes transcript --since 5m         # last 5 minutes
minutes transcript --status           # check if active

This lets you follow what's being discussed mid-meeting. The live output is rougher than the final transcript produced after stop -- it prioritizes speed over accuracy.

Check status:

minutes status

Returns JSON: {"recording": true, "pid": 12345} or {"recording": false}

What you get

A markdown file at ~/meetings/YYYY-MM-DD-title.md with:

• YAML frontmatter (title, date, duration, type)
• Timestamped transcript
• Summary, decisions, and action items (if LLM summarization is configured)

File permissions are set to 0600 (owner-only) because transcripts contain sensitive content.

First-time setup

If the user hasn't set up minutes before, they need a speech model:

Whisper (default):

minutes setup --model small

This downloads a ~466MB model. For faster but lower quality: --model tiny (75MB). For best quality: --model large-v3 (3.1GB).

Parakeet (opt-in, lower WER, fast on Apple Silicon):

minutes setup --parakeet                           # English (tdt-ctc-110m, ~220MB)
minutes setup --parakeet --parakeet-model tdt-600m  # Multilingual v3 (~1.2GB)

Requires parakeet.cpp installed. See docs/PARAKEET.md for full setup.

Gotchas

• "model not found" → Run minutes setup --model small (whisper) or minutes setup --parakeet (parakeet). This is the most common first-run error.
• "already recording" → Run minutes stop first, or minutes status to check. If the PID file is stale (process crashed), minutes stop will clean it up.
• No audio captured / empty transcript → Check that the right input device is selected in System Settings > Sound. On MacBooks, the default mic works for in-person conversations but won't capture system audio.
• For Zoom/Meet/Teams audio → You need BlackHole to capture system audio. See references/audio-devices.md in this skill folder for the full setup guide.
• Recording runs but transcription is garbage → The tiny model is fast but low quality. Upgrade to small or medium for real meetings: minutes setup --model small.
• "permission denied" on output file → Output files are 0600 (owner-only). This is intentional — transcripts contain sensitive content. Don't chmod them to be world-readable.
• Long meetings (>2 hours) → Transcription time scales with duration. A 2-hour meeting with the small model takes ~3-5 minutes on Apple Silicon. The tiny model is ~4x faster but much less accurate.
• Recording process disappeared → If you close the terminal tab where minutes record is running, the recording stops but may not process. Always use minutes stop from another terminal.