메뉴
Talk to Claude by voice. Records the mic, transcribes with ElevenLabs Scribe (hosted STT, no local Whisper), sends the text to Claude via the local `claude` CLI, and speaks the reply back with ElevenLabs TTS. Use when the user wants a spoken/voice conversation, voice input, or to dictate to Claude.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Read and manage Zergboard cards, boards, and workspaces. Use when the user asks to check Zergboard, view cards, find tasks, check sprint/cycle status, manage work items, invite guests, or interact with their Zergboard boards from the CLI.
Blink Camera Skill
Order food on DoorDash for the Beck family using the DoorDash CLI (dd-cli), composing a balanced order that covers everyone's tastes and is age-appropriate for the kids, then placing it only after Idan approves the exact items, total cost, tip, card, and ETA. Use when Idan wants to order dinner/food for the family, wants a DoorDash recommendation, or says "order us food." Reads Personal/Family Food Profile.md for tastes, allergies, and rules.
Create, edit, publish, schedule, and manage Substack posts. Use when the user asks to draft or publish a Substack post, list/edit their Substack drafts, schedule a post, or upload images to Substack. Supports multiple accounts/publications.
Generate audio or video of a specific person (the user) saying arbitrary text. Clones the user's voice with ElevenLabs (Professional Voice Clone) and drives a talking-head video with fal.ai audio-driven avatar models. Use when the user wants to make their voice say something, produce a talking-head video of themselves, or build/manage their voice/face twin.
Send / track / download DocuSign envelopes from the CLI. Anchor-string-based field auto-placement (no manual tab dragging). JWT Grant auth — one-time browser consent, then any number of programmatic sends. Use when the user wants to send a PDF for signature, check envelope status, list recent envelopes, or bulk-send to multiple recipients.
| name | voice-mode |
| description | Talk to Claude by voice. Records the mic, transcribes with ElevenLabs Scribe (hosted STT, no local Whisper), sends the text to Claude via the local `claude` CLI, and speaks the reply back with ElevenLabs TTS. Use when the user wants a spoken/voice conversation, voice input, or to dictate to Claude. |
A hands-free voice loop for Claude:
mic → ElevenLabs Scribe (STT) → Claude (local `claude` CLI) → ElevenLabs (TTS) → speaker
Why not Whisper / "use Claude to transcribe"? Claude has no audio input — it can't transcribe. So a fast hosted speech model (ElevenLabs Scribe) does the STT, and Claude is the brain in the middle. This avoids running Whisper locally (slower, heavier).
Reuses the ElevenLabs API key already set for eleven-labs-skill
(~/.claude/skills/eleven-labs-skill/config.json, or ELEVENLABS_API_KEY).
The brain uses your existing Claude Code login via the claude CLI — no API key needed.
pip3 install -r ~/.claude/skills/voice-mode/requirements.txt # sounddevice, numpy, requests
python3 ~/.claude/skills/voice-mode/voice_mode.py setup # verifies ElevenLabs + claude CLI
macOS will prompt for microphone permission the first time the terminal records.
The sounddevice wheel bundles PortAudio — no brew install needed.
# Continuous conversation (default). Speak, pause, hear the reply. Say "goodbye" or Ctrl-C to stop.
python3 ~/.claude/skills/voice-mode/voice_mode.py converse
# One turn: record → answer → speak
python3 ~/.claude/skills/voice-mode/voice_mode.py once
# Utilities
python3 ~/.claude/skills/voice-mode/voice_mode.py listen # record → {"text": ...}
python3 ~/.claude/skills/voice-mode/voice_mode.py say "hello" # TTS only
python3 ~/.claude/skills/voice-mode/voice_mode.py stt clip.wav # transcribe a file
python3 ~/.claude/skills/voice-mode/voice_mode.py mics # list input devices
It auto-detects when you stop talking (silence-based cutoff), so there's no push-to-talk.
config.json)| key | default | notes |
|---|---|---|
voice_id | 21m00Tcm4TlvDq8ikWAM | ElevenLabs voice (Rachel, a standard voice) |
stt_model | scribe_v1 | ElevenLabs Scribe |
tts_model | eleven_turbo_v2_5 | low-latency; use eleven_flash_v2_5 for min latency |
brain | claude | claude (Claude Code CLI) or codex |
silence_secs | 1.2 | trailing silence that ends an utterance |
max_utterance_secs | 30 | hard cap per turn |
Change the voice: voice_mode.py setup --voice-id <ID> --voice-name "<name>"
(browse voices with the eleven-labs-skill). Idan already has a personal voice clone
from the digital-twin skill — pass its voice id here to have Claude reply in his voice.
Switch the brain to Codex: voice_mode.py setup --brain codex.
claude -p (fresh context per call — avoids the
--continue staleness issue), keeping the last ~6 exchanges.