george-blaze/srtc

SRTC — Screen Recording To Claude

SKILL_DIR is the directory containing this SKILL.md (normally ~/.claude/skills/srtc). Every command below runs from there. The tool is python3 "$SKILL_DIR/vtc.py"; the digest engine is python3 "$SKILL_DIR/screencast/screencast.py".

What it does

• ⌃⌘S toggles a screen + microphone recording (menu-bar agent vtc-hotkey). Esc cancels. 10-minute cap with a 20-second countdown sound + red panel.
• On stop: 1 fps sampling → unchanged seconds dropped → 2×2 timestamped sheets; mouse log → zoomed "pointer sheets" for circled / rested-on / clicked spots; Whisper (local, streaming while recording) → timestamped transcript.
• Delivery: brings Claude to the front and pastes images then the prompt into the open chat (user presses Enter).
• A rule in ~/.claude/CLAUDE.md (installed from claude-rule.md) makes the narration the request.

Commands

Ask	Run
status / is it recording	python3 "$SKILL_DIR/vtc.py" status
digest an existing recording (any .mov/.mp4)	python3 "$SKILL_DIR/screencast/screencast.py" digest "<file>" → then Read the printed DIGEST.md, then each sheet in sheets/ and pointer/
storage usage	python3 "$SKILL_DIR/vtc.py" usage
clean old takes / clean everything	python3 "$SKILL_DIR/vtc.py" clean / python3 "$SKILL_DIR/vtc.py" clean --all
change hotkey / mic / cap / retention	edit ~/.config/vtc.json (keys: hotkey, mic, cap_seconds, warn_seconds, keep_days, deliver, auto_send, asr), then launchctl kickstart -k gui/$(id -u)/com.srtc.hotkey
install / reinstall	"$SKILL_DIR/install.sh" (the user must run it: it registers a login agent and a keychain certificate)

When the user asks a question about a recording they just made, the digest folder is recordings/<date>.digest/ next to the video (default ~/Movies/SRTC/, or recordings in ~/.config/vtc.json): read DIGEST.md first, then the sheets it lists.

Troubleshooting (check in this order)

1. Hotkey dead, menu bar says "hotkey blocked: password field" → some app holds secure keyboard input (a focused password field, often a browser). Click out of it. ioreg -l -w 0 | grep -o '"kCGSSessionSecureInputPID"=[0-9]*' names the pid.
2. Hotkey dead, log says "not trusted" (~/Library/Application Support/VideoToClaude/hotkey.log) → Accessibility grant lost (happens after rebuilding without sign.sh). System Settings → Privacy & Security → Accessibility: remove vtc-hotkey (−), add it again (+ or drag the binary in). The agent arms itself within 2 s.
3. "Nothing was recorded" → Screen Recording or Microphone grant missing for vtc-hotkey; Screen Recording only applies after launchctl kickstart -k gui/$(id -u)/com.srtc.hotkey.
4. Transcript empty or garbage → check the digest's "audio peak / average" line; below −40 dB average means the wrong mic. mic in ~/.config/vtc.json is a priority list; python3 "$SKILL_DIR/screencast/screencast.py" devices lists devices.
5. Transcript OK but late → Whisper model missing (models/whisper-large-v3-turbo/weights.safetensors); see README. Logs: ~/Library/Application Support/VideoToClaude/{hotkey,runner,asr,recorder-ffmpeg}.log.

Never run install.sh, sign.sh or launchctl bootstrap yourself; show the command and let the user run it.