$digital-twin

A Claude Code skill · Seedream 5 Pro · Seedance 2.5 · ElevenLabs

Your digital twin

Photoreal stills of you in your own room, talking clips of you, and lines in your own voice, good enough that people who know you say “that’s you”.

The workflow comes from one creator’s twin: more than 160 numbered stills, over 20 talking drafts and a voice taken from 8.7 to 9.2 out of 10. Every rule came from a correction he made or a measured failure.

See the workflow
Mai, an example presenter, at her desk in the evening, talking into the lensStill 2 · evening
Evening, 85 mm · eye contact · the one picked for the clip
The same room at early sunrise, from slightly above the monitorStill 3
Early sunrise · 35 mm, slightly high
The same room in late afternoon light, three-quarter view past the lensStill 1
Late afternoon · 50 mm, 3/4 past the lens

In Fil

The workflow, step by step

Mai’s batch in Fil’s Digital twin workflow: her face frames and her room dropped in, three Seedream 5.0 Pro stills, the iMac fixed in one edit, her line on ElevenLabs v4, then still 2 as a Seedance 2.5 draft, made final at 1080p. Drawn live in HTML from one clock, with Fil’s own prices.

Mai is one of Fil’s example presenters, standing in for you; her room’s photo is the corner of it still 2 sees. Credits as Fil prices them: 12 a still, 9 an edit, 1 for the line, 55 for the draft and 289 to make it final.

What it makes

Three things that should pass for you

Your own eye is the final judge. Blind judges and scorers find the faults; you decide.

Mai at her desk in late afternoon lightStill 1

Stills in your real room

Three at a time. The room, gear and face stay fixed; only the camera and the time of day change.

model
Seedream 5 Pro, 16:9, 2k
price
$0.10–0.12 a still
gate
your pick; face measured against your real frames

Talking clips

Five or six seconds from a picked still, with a line in your voice. One draft per still, measured before you see it; the final, from the draft you like.

model
Seedance 2.5 draft 480p, then its final 1080p
price
$0.52–0.59 a draft · $3.12 a final
gate
words exact; mouth and head against how you talk
Mai’s line · Kristen voice, v4, read conversationally

Lines in your voice

An instant clone from three minutes of your cleanest speech, tuned with an audio tag and scored against your real voice.

model
ElevenLabs v4 + [calm]
price
~$0.01 a line
gate
≥ 9 on the scorer; your ear on a before/after

The workflow

Eight stages, each with a gate

Price every run before it starts, log every spend, and put every result on one canvas with its recipe and your verdict.

  1. 0

    Consent and setup

    Only your own face and voice, or someone who gave documented consent. Accounts, a project folder, a canvas, a ledger.

    NeedsFil, the Higgsfield CLI, ElevenLabs, ffmpeg, whisper, Swift
    Gateconsent recorded; price before every spend
  2. 1

    References from your recordings

    Face crops from your own talking-head videos: a lean set of 8 (no smiling or mid-word frames), angle sets, 3 room photos, 5 held-out frames for the judge.

    Costfree · extract_frames.sh, faces.swift, measure.py
    Gatenever a selfie, a generated face or someone else’s frame
  3. 2

    The identity block

    One paragraph from a blind judge’s feature list. One fix, one phrase; say the near misses (“not a quiff”). Longer blocks over-correct.

    Costfree
    Gateunder 4,000 characters with the rest of the stack on Fil
  4. 3

    Stills, three at a time

    Identity → shared workspace → shot → gear and finish. Your real room with its photos last; only the camera and the time change.

    CostSeedream 5 Pro: 12 Fil ($0.12) or 2.5 Higgsfield (~$0.10)
    Gateyour pick; chin and width within ±0.05 n of your real frames
  5. 4

    Audit and fix

    Two fresh reviewers in 3×3 tiles. One edit lists every mistake; only the flagged regions are pasted back, with the face masked.

    Costan edit: 9 Fil or 2.5 Higgsfield · fix_still.py
    Gateno wrong gear, grey hair, extra objects, stands or mics
  6. 5

    Your voice

    An instant clone first. v4 with an audio tag if the pitch runs high. Score 7–10 s lines; then a before/after for your ear.

    CostElevenLabs characters · score_voice.py is free
    Gate≥ 9 out of 10; “sounds like me”
  7. 6

    Talking clips

    Seedance 2.5 draft from an audited close-up, the still as image reference too, your line as the audio reference, the soft-mouth clause.

    Cost15 Higgsfield (~$0.59) or ~52 Fil ($0.52)
    Gateone draft per still; English exact; mouth p50 near yours
  8. 7

    Deliver

    A canvas section per batch, notes on every node, picks marked and moved to a keep section, one question: which are you?

    Costfree
    Gateledger and log updated before you report

Worked example

One batch, start to finish

Mai is one of Fil’s example presenters, an AI portrait standing in for you. Eight frames of her face from one of her clips went in as references, one room described in words, and three shots that change only the camera and the time of day. Every picture, sound and clip here was made that way.

Eight frames of Mai’s face from one of her clipsreferences
Her face · eight frames from her own video, never a selfie
Still 1 as it came: an iMac with a chin and a logo on the deskStill 1 · as it came
The most common miss · all three drew an iMac, chin and logo, though the prompt said no chin
Still 1 fixed: a thin display with no chinStill 1 · fixed
Fixed · one Seedream edit, then only the display pasted back (drift 5.3 levels outside it)
Still 2, evening, eye contactStill 2
Evening, 85 mm, eye contact · picked for the talking clip

Draft · 480p · 18 Higgsfield credits

Final · the same take, 1080p · 96 credits

Talking clip · Seedance 2.5, 6 s, from still 2, with her line in a stock ElevenLabs voice as the audio reference: the draft, then the final made from it

Checked before anyone sees it

Written down by whisper, the words came back exactly: “This course is different. The agent writes the code and you shape how it feels.” Read conversationally at 175 words a minute, it fills the clip. The face, the display and the room held through the clip, and the final keeps the draft’s take with the detail back.

3 + 3stills, then one edit each
~60Fil credits for the batch and fixes
18 + 96Higgsfield credits: draft, then 1080p final
exactthe line, word for word

Voice

A score you can trust, then your ear

A speaker model (ECAPA-TDNN) compares each clip with your own clips, after the same band-pass and loudness chain, so a different mic can’t pass for a different person.

score = 0.7 × similarity + 0.3 × pitch
similarity: your held-out clips = 10, other people = 0
pitch: 10 inside your own p10–p90 range
  • 9 is you. One creator’s own voice, denoised, scored 9.3.
  • Score 7–10 s lines. A 4 s clip of the right voice scored 4.7.
  • Stability didn’t move v4’s pitch. The tag did, and whisper confirms it isn’t spoken.
  • Professional voice clones have a trap: adding samples resets verification, and the voice goes offline everywhere. You get five tries, then a 24-hour lock. Keep an instant clone as the stand-in.

One creator’s clone on each model and tag, mean of three test lines

The dashed line is 9, his own processed voice. The amber bar is the one he kept.

What we learned

Ten rules, and the failure each one prevents

Use a lean reference set

Preventsa face 8–10 % long in the chin and 7 % wide. Seedream copies its references’ proportions, so smiling and mid-word frames stretch every image.

Your real room, only the camera changes

Preventsinvented lofts and cabins that read as fake, break continuity and, in that project, ran long in the chin.

Three at a time, numbered

Preventsbig batches nobody can judge. Three numbered shots make a pick one line: “144 145 are good”.

Angle references for every angle

Preventseye-contact and side shots drifting into someone else. Front frames gave the closest faces of all.

Reference videos as words only

Preventsanother person’s face blending into yours. Describe the composition and the light; never wire their frame in.

Words become features

Preventssilver hair from “grey scalp”, blue hair from “blue-black sheen”, a longer chin from “short chin”. One fix, one phrase.

Audit before video

Preventsan iMac, a ghost finger or a second laptop showing in every frame of a clip. Fix with one edit and paste back only those regions.

Calm mouth, calm head, one draft per still

PreventsSeedance’s 50 % too-wide mouth, a busy head, and near-duplicate drafts that waste money.

An instant clone and an audio tag

Preventsa clone that has your timbre but speaks 30 Hz high. [calm] on v4 scored 9.2; stability changed nothing.

Never add samples to a voice in use

Preventsa professional clone going offline on every model, then a 24-hour verification lock.

Numbers that shipped

One person’s measured results

Measure your own. These show the scale and the targets the workflow aims at.

WhatValue
his face, real frameschin 0.80 n · width 2.63 n (nose) · 2.24 n (mouth)
a good takewithin ±0.05 n of those; his room plus front refs gave 0.78–0.83 on most shots
blind identity plateau8.75 / 10 · judges drift 0.4–1.25 between rounds
his mouth when talkingp50 0.092 · Seedance alone 0.135 · best drafts 0.085–0.10
his head when talkingturn sd 0.100 · tilt 3.9° · Seedance alone 0.034 · 1.7°
his voicemedian 116 Hz · p10–p90 106–131 Hz · clone 9.2 / 10
what the whole project spentabout $45 over three days (~1,300 Fil + ~800 Higgsfield credits)

Use it

Start with your recordings

Put the skill in ~/.claude/skills/digital-twin/, then ask in plain words.

Minimal promptUse $digital-twin to build my digital twin from my own recordings in ~/Videos: pick a lean reference set and held-out judge frames, write my identity block, make 3 Seedream 5 Pro stills in my real room with only the camera changing, measure and audit them, clone my voice and score it to 9/10, then make one Seedance 2.5 draft per picked still. Price every step first.
Just the voiceUse $digital-twin to clone my voice from my last three recordings: label my own speech, build an instant clone from my cleanest 3 minutes, try v4 with and without audio tags, score every take with the ECAPA scorer on 7–10 s lines, and send me a before/after with my real voice.

What’s in the skill

  • SKILL.md: the stages, the rules, what fails and the fix, numbers, a checklist.
  • references/: consent, references, identity template, stills, workspace, cinematic looks, audit, voice, speaking, video, judging, costs, the Fil workflow.
  • scripts/: frames and face crops, landmarks and proportions, a round builder and batch runner, numbered sheets, the paste-back fix, film grades, speaking stats, draft measuring.
  • scripts/voice/: speaker labelling, dataset scan, instant-clone builder, a line generator, the scorer, a training watcher, an API helper that never prints your key.