Live Recording Skill | Agent Skills

Live Recording

Always-on environmental audio capture. Records continuously, transcribes with whisper-cpp (Metal-accelerated), appends timestamped entries to an org file viewable live in Emacs.

Activation

1. Start tmux capture session

tmux new-session -d -s capture \
  'while true; do bash ~/v/scripts/always-on-capture.sh; echo "[restart] $(date)"; sleep 2; done'

The auto-restart loop ensures crash resilience. The script records 6s WAV chunks via ffmpeg avfoundation device :1, silence-detects via sox RMS (< 0.005), transcribes non-silent chunks with whisper-cli --model ~/v/ggml-base.en.bin, and appends ** [timestamp]\ntext\n entries to ~/v/audio-capture.org.

2. Open in Emacs with live updates

M-x server-start                          ;; required for emacsclient
C-x C-f ~/v/audio-capture.org             ;; open the file
M-x auto-revert-tail-mode                 ;; live tail as entries append

auto-revert-tail-mode (not auto-revert-mode) tails the end of the buffer as new entries arrive, keeping your cursor at the latest transcription.

3. Monitor (optional)

tmux attach -t capture          # watch capture output
tail -f ~/v/audio-capture.org   # terminal tail

Stopping

tmux kill-session -t capture

Configuration (in always-on-capture.sh)

| Variable | Default | Purpose | |----------|---------|---------| | AUDIO_DEVICE | :1 | avfoundation device (:0 = screen, :1 = mic) | | CHUNK_SECONDS | 6 | Recording window per chunk | | SILENCE_THRESHOLD | 0.005 | RMS below this = silence (suppressed) | | MIN_TEXT_LEN | 6 | Discard transcriptions shorter than this | | MODEL | ~/v/ggml-base.en.bin | Whisper GGML model path | | ORG_FILE | ~/v/audio-capture.org | Output org file |

Troubleshooting

| Symptom | Cause | Fix | |---------|-------|-----| | All chunks "no audio" | Wrong device index | ffmpeg -f avfoundation -list_devices true -i "" to find mic | | All chunks "silence" | Threshold too high | Lower SILENCE_THRESHOLD in script (default 0.005) | | Emacs not updating | No server-start | M-x server-start then reopen file | | Emacs not tailing | Wrong mode | Use auto-revert-tail-mode not auto-revert-mode | | Script exits immediately | set -e + non-fatal error | Script uses set -uo pipefail (no -e) |

Related Files

~/v/audio-capture.org -- live output file
~/v/scripts/always-on-capture.sh -- capture script
~/v/audio_acset.duckdb -- structured audio database (ACSets schema)
~/v/scripts/audio-capture-org.py -- Python/mlx-whisper alternative

Agent Skills: Live Recording

Install this agent skill to your local

Skill Files