Agent Skills: Digital Twin

Digital Twin

Create a voice + video twin of yourself, then make it say arbitrary text.

Vendor split: ElevenLabs for the cloned voice (best fidelity), fal.ai for the audio-driven talking-head video. Keys are reused from eleven-labs-skill and fal-video-skill configs automatically, so you usually don't re-enter them.

Pipeline

voice (PVC) ──────────────────────────────►  say  = cloned-voice audio   (audio twin)

reference photo ─► shots (nano-banana) ─► pick a keyframe ─┐
voice (PVC) ───────────────────────────────────────────────┴─► video = kling talking-head mp4

Setup

pip3 install elevenlabs requests fal-client   # ffmpeg recommended; nano-banana-pro skill for `shots`
python3 ~/.claude/skills/digital-twin/digital_twin.py status
# Optionally store keys directly (else inherited from the other skills):
python3 ~/.claude/skills/digital-twin/digital_twin.py setup --elevenlabs-key KEY --fal-key KEY
# Keep media out of the skill repo (recommended) — store it anywhere:
python3 ~/.claude/skills/digital-twin/digital_twin.py setup --data-dir ~/Documents/DigitalTwin

1. Clone the voice (Professional Voice Clone)

PVC gives near-indistinguishable fidelity. It wants 30+ min of clean solo audio (≈3 hrs ideal), needs a Creator+ ElevenLabs tier with a free professional-voice slot, and includes a human verification step (you record yourself reading a captcha).

# 1. create the voice + upload samples (mp3/wav/m4a)
python3 .../digital_twin.py voice-create "Idan" sample1.mp3 sample2.mp3 ...

# 2. get the captcha text to read aloud
python3 .../digital_twin.py voice-verify
# ...record yourself reading it, then submit the recording:
python3 .../digital_twin.py voice-verify --recording captcha.mp3
#   (or manual ID path:  voice-verify --id-file consent.pdf)

# 3. train, then poll until done (training takes hours)
python3 .../digital_twin.py voice-train
python3 .../digital_twin.py voice-status

The resulting voice_id is stored in config.json and used by say/video.

Prototype today without waiting for PVC — Instant Voice Clone (1–5 min audio, immediate):

python3 .../digital_twin.py voice-instant "Idan" short_sample.mp3

Point at an existing voice / list voices:

python3 .../digital_twin.py voices
python3 .../digital_twin.py use-voice VOICE_ID

2. Enroll a face (keyframe still)

# Use a photo as-is (recommended starting point):
python3 .../digital_twin.py enroll-face "/path/to/photo.png"

# Or synthesize a clean front-facing headshot from reference photos (nano-banana-pro):
python3 .../digital_twin.py enroll-face photo1.png photo2.png --generate

A short (15–30s) video clip of the subject yields materially better likeness than a still, but every avatar model here works from a single photo.

2b. Generate reference shots (optional — art direction)

Turn your real reference frame(s) into a library of styled keyframes — different expressions, angles, wardrobe, or settings — via nano-banana-pro, with identity anchored by the reference image(s). Use real, current-look frames as references for consistency.

# default set: headshot, slight smile, 3/4 angle, blazer-in-office, casual home-office
python3 .../digital_twin.py shots ref1.jpg ref2.jpg

# custom directions
python3 .../digital_twin.py shots ref1.jpg --shots "in a suit on a conference stage" "outdoors, golden hour"

Frames land in data_dir/keyframes/shots/. Drive any one with the video stage:

python3 .../digital_twin.py video "..." --keyframe <shots/...png>

3. Generate

# Audio-only twin
python3 .../digital_twin.py say "Hey, this is Idan." --out hello.mp3
python3 .../digital_twin.py say --file script.txt

# Full talking-head video
python3 .../digital_twin.py video "Hey, this is Idan." --out hello.mp4
python3 .../digital_twin.py video --file script.txt --model infinitetalk

Avatar models (--model)

| key | fal endpoint | notes | |-----|--------------|-------| | kling (default) | fal-ai/kling-video/ai-avatar/v2/pro | most expressive; 1080p; up to ~5 min | | omnihuman | fal-ai/bytedance/omnihuman/v1.5 | turnkey single-photo; ~30s audio cap/call | | infinitetalk | fal-ai/infinitalk | long-form, strong identity, effectively unlimited |

python3 .../digital_twin.py models lists avatar + TTS models.

Notes & caveats

• Long scripts: OmniHuman caps near 30s of audio per call; the skill warns and you should switch to --model infinitetalk (or kling) for monologues.
• fal input fields: avatar models are sent image_url + audio_url + prompt (verified against the live fal OpenAPI schemas; infinitalk requires prompt, the others default it). fal slugs and input names evolve; if a model errors on validation, check its page on fal.ai and adjust AVATAR_MODELS / the payload in digital_twin.py.
• TTS model: defaults to eleven_multilingual_v2. eleven_v3 is more expressive but may be gated on your account; eleven_flash_v2_5 is low-latency.
• Motion prompt: video sends a default webcam-style prompt (static, locked-off, no zoom / Ken Burns). Override per call with --prompt "..." or set default_video_prompt in config. This matters: without it, avatar models tend to add a slow push-in.
• Storage: set data_dir (setup --data-dir DIR) to keep media out of the skill repo (which is .gitignore'd and pushed to GitHub). Outputs → data_dir/output, keyframes → data_dir/keyframes, shots → data_dir/keyframes/shots. Defaults to the skill dir if unset.
• Consent & safety: only clone your own likeness. A voice+face twin is high impersonation risk — keep the cloned voice_id, keyframe, and outputs access-controlled and never expose them to third parties via bridges/CRM without explicit gating.

#voice #video #avatar #elevenlabs #fal #digitaltwin