format-ai-actor-demo: she calls it out, the real band proves it

1. Read first

• knowledge/ad-formats/_FORMAT-CORE.md (doctrine, lane limits, gates, launch). Binding. Photoreal people run on ugc-omni's G0-G4 gates (§10 here).
• knowledge/ad-formats/_MODEL-REGISTRY.md. Every model below is named by ROLE; the registry says which model fills it today.
• references/beat-sheet.md (skeleton, word budget, line rules, register cells), references/sources.md (davidaistar videos step by step, what we keep and drop, our receipts).
• skills/ugc-omni/SKILL.md + scripts/ugc_omni/omni.py: THE photoreal production engine (Fish 10-06). This skill hands it a job.json; it never re-specifies the engine. Another session owns both: reference, never edit (same for skills/ugc-omni/references/davidaistar-diff-2026-10-06.md). Example job for this format: references/omni-job.example.json.
• output/dawn-combo-map-2026-10-05/MAP.md §2B (recipe B) and knowledge/brands/dawnbands/product-truth.md, knowledge/brands/dawnbands/claims-register.md.
• Memory: feedback_ai_avatar_broll_lane.md (the 9-30 rules ugc-omni's gates implement), project_dawn_yapper_2026-10-04.md (why ungated band-in-hand failed), feedback_reference_images.md, feedback_ugc_avatar_wardrobe.md, corrections-hot.md (10-05: simplest recipe first; beat the leader, don't mirror it).

2. When to use / when not

• Use when Fish wants a short, cheap, high-volume video cell: one callout line that stops her, the band shown doing its one job, and a guarantee. It's also the right cell for testing a winning static's callout as video (same stake, one variable: format).
• Evidence status: untested on Dawn.
• Coded Dawn data leans against what this format does by default. In statics, device: demo ran $153 at ROAS 1.10 (n1), testimonial $26.8K at 1.50, callout-direct $190K at 1.86. Short cartoon VO ran $314 at ROAS 0 (n10). MAP §2B: a person reads like a testimonial, and testimonials lose.
• So the actor's job is to deliver a recipe-B callout, not a review.
• Outside Dawn, the source claims are GMV claims (shapewear clip ~$11K GMV, sleep-gummy challenge ad ~$68.8K). Those are TikTok Shop affiliate numbers, not Meta CPA.
• Don't use for:
• The accused-mom story (recipe A → format-song / format-spoken-drama).
• Anything over ~35 s (→ ugc-omni long form).
• Any generated product frame that has not passed ugc-omni's product gate (pick --product ≥7/10, right display state). Ungated band-in-hand scored 3-4/10 on 10-04.
• A clone of a competitor's ad (10-05 rule: competitors are the gap map, not the template).
• Any paid step before G0 (Fish approves script + hooks + storyboard).

3. The format

• 9:16, 15-35 s (target 25-30 s), 6-9 spoken lines. One AI actor (a mom 35-54 who already SOLVED it) talks to a propped or held phone in one home location. There's a scene change at the demo.
• What the viewer sees:
• Hook to camera.
• One problem line, over real "kid sleeps through it" b-roll.
• The demo: on the scene change the band is already lying on the counter in front of her; on the demo line she picks it up and rests it across her palm toward the lens (continuous reveal, gated frames). Then b-roll of the band doing its job (on a sleeping kid's wrist, buzzing, the kid sitting up): gated omni b-roll, or real footage as the fallback. Her voice keeps running underneath.
• The result back on her face (calm kitchen, coffee).
• CTA, then an end card (optional, --end-card; off by default per Fish 10-06) with the real product cutout.
• Captions are script words on every line, plus a hook overlay on frame 0. "Dramatization" label on-screen or in primary text (Fish 10-06, all photoreal people ads).
• Reference shapes, paraphrased (never copied):
• davidaistar Rqim7SlMtZg: AI actress, proven TikTok Shop motion, product kept real.
• gpzy2ofG0Dw: "playful challenge" hook → time jump → proof.
• PmXqgiFn9_Y: 15-30 s "caveman English" native-voice talking head.
• CZwfTy7jZcQ: hook → problem → product → demo b-roll → CTA.
• Register options (one per ad): plain callout (default), challenge, caveman cell (Fish's ruling needed, see §11).

4. Why it sells

• Recipe B (MAP §2B):
• Brand-voice-style callout to HER ("If you're still the one waking him up…").
• Her daily burden as the stake.
• Exhaustion → relief.
• Authority = guarantee or science, not the actor's opinion.
• Payoff = his independence or her quiet morning.
• The actor adds a human face to the callout. The real demo adds proof the viewer can't get from a static: the band visibly buzzes and a kid gets up.
• Problem-aware: line 1 names her by an observable symptom ("…the seventh alarm", "…still knocking at 6:40"). The hook = emotion + curiosity gap + high stakes, starts in action, zero backstory.
• Why the product gate: a demo only persuades if the product looks real. A morphing generated band undoes the proof and breaks product truth, so every frame that carries the band is scored against the real pack before it becomes video (ugc-omni §5); real footage covers any shot that won't pass.

5. Beat sheet (full rules + word budget in references/beat-sheet.md)

Budget for 25-30 s at the locked Omni pace (~2.6-3.0 wps): 65-85 words, every line ≤10 s (Omni's cap), most lines 8-16 words.

6. Pack spec (ugc-omni job.json, one job per ad)

• Step 0 — combo plan (feedback loop, workflow-combo-loop): python3 scripts/combo_loop/plan.py --format format-ai-actor-demo --brand dawn → pick one combo (a paying parent + ONE changed dimension) and paste its block into the pack front matter / job or spec JSON ("combo": {...}; no file → tags.py register --prefix): combo_avatar combo_angle combo_pov combo_authority combo_stake combo_stake_on combo_emotion combo_root_cause combo_mechanism combo_payoff combo_device combo_format combo_parent combo_variable combo_ad_prefix (= the launched Meta ad-name prefix). python3 scripts/combo_loop/tags.py check <pack> must PASS before concept approval; the weekly refresh reads results back by that prefix.

• Start from references/omni-job.example.json (this format's shape, illustrative lines), validated against skills/ugc-omni/job.example.json's schema. New slug per run.

• Core §5 keys go in as top-level fields (omni.py ignores extra keys): format: ugc-omni, format_skill: format-ai-actor-demo, brand, created, status, parent (the paying static/native whose callout this carries, with its CPA), iteration (the ONE variable), title, overlay (hook text), label: "Dramatization".
• Engine keys: style: photoreal, mode: cuts, aroll_model: flash, resolution: 1080p, image_arms: ["gpt2", "sd5"] (Nano Banana retired 10-06), product_refs: ["@dawn"], product_name: "Dawn Band", product_truth = core §1 wording + ugc-omni's display-off sentence. say: {"\\bADHD\\b": "A-D-H-D"}; numbers spelled out in text.
• lines[] from the beat sheet: id, text, visual ("AI UGC" or "broll:<id>"), hook (H1-H3 on line 1 only), frame (avatar until the demo, scene2 from the demo on), plus src (the VOC/stake ID; extra key, required by this skill).
• Product line (beat 3, first demo line): visual: "AI UGC", frame: "scene2" (scene2 is picked --product: band lying on the counter), product: true, end_frame: "holding".
• Every other A-roll line: no product key; hands low.
• prompt (the lock): scene starts with the camera ("Static locked-off shot, UGC iPhone footage of a mom in her kitchen in the morning..."), says: "The woman says:", rules ("Pace... One take, no jump cuts. Same avatar, same location, same camera framing. Do not recompose the shot. No captions, subtitles or on-screen text."), and product_action: one fixed sentence, natural hold: "She picks the band up off the counter and rests it across her open palm toward the camera, close to the lens; it stays in her palm and in frame." Never pinched upright between fingers.
• avatar: genetic change, wardrobe: "keep" (Fish 10-06; if the ref's top is grey/blank, pick another ref or name a specific real garment, never plain/basic; no grey tops), accessories, audio logic, background: keep. No watch or bracelet on her wrists (reads as a smartwatch).
• broll: {prompt, outfit, motion, duration} per KJ; band shots add product: true (display off); scenes without her set avatar: false. Real-footage fallback: same id, prompt: "REAL FOOTAGE", avatar: false; copy the clip to <job>/broll/<id>.mp4 and never run broll on that id.
• Voice: no voice preset unless lock takes drift (then voice --spend, one change per roll).

7. Visual grammar

• A-roll: chest-up, phone propped at eye level or held. Never more than ~6 s straight on her face (board warns at 7 s); never 4+ b-roll lines without a cut back to her. Hands low on every non-product line.
• Scene change (scene2: same actor, new spot, e.g. kitchen counter) lands on the demo line, with the band already in shot. That's KJ's ~60% re-hook pulled earlier for a short ad, and where he lands the product.
• Product on screen (ugc-omni §5 gate, Fish 10-06): CLOSE framing (in a wide shot the band loses its ribs). Display off: every product prompt carries "Display off: a plain black ribbed silicone strap of even width all the way around, no visible screen, no screen pod, no raised module, no glossy rectangle; it looks like a plain sports band." Worn on a wrist or resting across a palm, never pinched upright. Continuous reveal: scene2 (band on the counter, --product) → product line → end_frame: holding (--product). product_action locked separately at G2.
• Band b-roll: gated omni b-roll (band on a sleeping kid's wrist; kid sits up) with pick --name broll_<id> --product. Real footage fallback when a frame won't pass: Fish's Dawn b-roll, python3 scripts/clip_library.py search --tags dawn-raw,product,waking-up --all -n 20 (folder clip-library/raw/dawn-broll-2026-09-10/DAWN BROLL/: Jessica kid sleeping wearing the band, band vibrating on a counter, Lori IMG_3496/IMG_3546 kid sits up, IMG_3458 wrist-POV alarm off, Kris IMG_3223-3234 teen out the door). Use .h264-clean.mp4 files, never the "captions" versions. Dark display preferred; a real lit LED only with Fish's ok.
• No readable text except overlay and captions. No logos. A boring, lived-in real reference frame, not an aspirational one (feedback_reference_images.md).

8. Audio

• A-roll voice = Omni's native voice from the locked prompt. No ElevenLabs on the actor (Fish 10-04). CAPITALISE the stressed word; captions drop the caps.
• Her voice runs continuously under the b-roll (assemble maps only A-roll audio, so real clips are muted).
• No music bed, or a very low one; a TikTok-Shop demo reads as a person talking.
• Pace = the locked take's native speed. verify flags <2.6 wps, holds ≥0.4 s, cut-offs, transcript <0.90.

9. Production (ugc-omni; every paid step dry-runs until --spend)

# run from ~/.openclaw/workspace; job.json per §6
O=scripts/ugc_omni/omni.py; J=output/ugc_omni/<slug>
# G0 (free): script + hooks + storyboard
python3 $O mine <real mom-talking-to-phone URLs...> --keywords alarm,morning   # FB/IG creators for 35-54
python3 $O board --job $J          # structure clean; only "frame not picked yet" errors may remain → Fish
# G1: reference frame + avatar + product frames
python3 $O refscan --job $J --src <chosen video>
python3 $O ref --job $J --pick $J/ref/<cand>.jpg
python3 $O avatar --job $J --n 3 --arms gpt2,sd5 [--spend]
python3 $O pick --job $J --name avatar --file $J/avatar/<pick>.png
python3 $O variant --job $J --name scene2 --product --arms gpt2,sd5 --prompt "<same woman at her kitchen counter, the band lying flat on the counter in front of her. Display off: ...>" [--spend]
python3 $O pick --job $J --name scene2 --product --file $J/frames/<pick>.png      # refuses <7/10 or wrong display
python3 $O variant --job $J --name holding --product --arms gpt2,sd5 --prompt "<same woman, same counter, the band resting across her open palm toward the camera, close. Display off: ...>" [--spend]
python3 $O pick --job $J --name holding --product --file $J/frames/<pick>.png
python3 $O board --job $J          # now fully clean
# G2: lock (motion proof): same line good 3x, hook first, then the product line
python3 $O lock --job $J --line H1-1 --n 3 [--spend]
python3 $O lock --job $J --approve $J/lock/<take>.mp4
python3 $O lock --job $J --line D-1 --n 3 [--spend]                 # product_action locked separately
python3 $O lock --job $J --approve $J/lock/<take>.mp4
# G3: every A-roll line + b-roll + product check
python3 $O gen --job $J --lines all [--spend]
python3 $O verify --job $J
python3 $O broll --job $J --id wristdemo --stage frame --arms gpt2,sd5 [--spend]
python3 $O pick --job $J --name broll_wristdemo --product --file $J/broll/<pick>.png
python3 $O broll --job $J --id wristdemo --stage anim [--spend]
# (fallback) cp "<real clip>.h264-clean.mp4" $J/broll/wristdemo.mp4
# G4: assembled cut
python3 $O assemble --job $J --hook H1                               # repeat H2, H3
python3 $O cost --job $J
python3 scripts/meta_preflight.py $J/final/<slug>-H1.mp4

• Cost (kie, omni.py): Omni A-roll/b-roll buckets $0.315 / $0.42 / $0.525 / $0.63 for 4 / 6 / 8 / 10 s (720p = 1080p price). Images ~$0.05 (gpt2) - $0.075 (sd5) each. Typical ad: avatar 3×2 arms + 3 variants×2 ≈ $1.1; 2 lock rounds ≈ $2.5-3.2; 6 lines + 2 hook takes ≈ $3.4; 2 b-roll ≈ $0.9; regens ~+30% → ≈ $9-11 per ad with 3 hooks. omni.py cost prices the job (it counts real-footage b-roll as Omni clips: subtract those). Check kie credit first (GET https://api.kie.ai/api/v1/chat/credit).
• No draft→final step: the G2 3× lock is the proof; the approved prompt makes the finals. Tiers: every A-roll take is Omni; tiers only pick b-roll (A = band/hands → gated omni or real footage; C = band-free scene → cheapest omni 4 s or a real clip).
• Image arms: pass --arms gpt2,sd5 (omni.py's own default still lists nbp; its owner changes that).
• Time: ref + avatar ~1.5 h by eye; product frames ~30 min; locks ~45 min; gen + assemble ~30 min.

10. Gates & QA checklist (ugc-omni G0-G4; every gate is Fish's)

• [ ] G0 script + hooks + board: parent named with CPA, ONE variable; every line has a src; stake caused by the mornings and lands on her; 2nd-person callout, not a review; mechanism lines are Fish's ruled lines; copy through direct-response-copywriter → humanizer → red-team (core §1). No paid step before G0.
• [ ] G1 reference frame (real creator video, refscan + eye) + avatar (genetic change, not the ref person; real clothes, no grey top; bare wrists) + product frames (pick --product passed, display off, natural hold, close).
• [ ] G2 lock: hook line, same take good 3× (the 9-30 motion proof), then the product line with product_action good 3×. Judge in a fixed order: face/hair/wardrobe stable → hands → line complete, in sync → room continuity → camera calm → clean ending → captions can sit clear of her mouth. One change per roll. No gen before G2.
• [ ] G3 every A-roll line watched (verify clean or each flag accepted; voice consistent by ear); every b-roll frame + band checked by eye against the real pack.
• [ ] G4 assembled cut by eye and ear: captions = script, "Dramatization" label, end card only if --end-card, meta_preflight run. Say what you could not verify (lip-sync, voice likeness). Scores never pass a gate alone; Gemini QA is not a gate.

11. Variants & test plan

• Minimal test (G1-G2, ~$2): avatar rolls + lock --n 3 on the hook line. Fish judges realism, lips and voice. Stop here if it fails.
• Full test: one ad, 3 hooks on a shared body (assemble --hook Hn). Hook directions: (a) her burden count ("the seventh alarm in this house is you"); (b) stake on her (the school letter / late to her own job); (c) challenge ("can a wristband get a kid who sleeps through six alarms up in one buzz?").
• Launch: own ad set in Testing CBO (core §6). Read at 7 days on CPA vs the parent static, plus 3-s hold and lifetime frequency. Never pause; reduce.
• Then one variable per ad set: register cell (plain / challenge / caveman); actor (age, ethnicity); setting (kitchen / car at drop-off / hallway); length (15 vs 30 s); band b-roll (gated omni vs real footage); lane (AI actor vs real creator fallback A).
• Caveman cell: broken, literal short sentences ("Alarm loud. Kid sleep. Mom tired."). It fails the said-out-loud rule on purpose. Only if Fish rules the register in, never on an ADHD line, native voice only.
• Side arms (one variable vs the locked omni baseline, ugc-omni §6 recipe discipline): mode: oneshot yapper single take (no b-roll; reveal via end_frame); H3 native-voice yapper take (yapper.py); avatar arm A/B (pick the other arm's roll as avatar_sd5, then lock --frame avatar_sd5). Seedance 2.5 product shots stay rejected (Fish 10-06).

12. Failure modes

• Band wrong in a generated frame (smartwatch pod, smooth ribs, lit face) → re-roll the frame (never patch); pick --product refuses <7/10; tighten framing and keep the display-off sentence; still failing → real footage.
• Hand drops / band leaves frame mid-line → product_action missing or unlocked; lock it on the product line (10-06 test: hand dropped at 0.25 s without it).
• Pinched-upright band → rewrite the hold (worn, or resting across the palm).
• Watch on her wrist → bare-wrist avatar; re-roll the avatar.
• Testimonial drift ("I love it, five stars") → rewrite to a callout plus guarantee/science authority (MAP §2B).
• Voice drifts between takes → regen the worst take; then the voice preset (one change); never ElevenLabs over her.
• Model says quoted words twice / adds words → transcript <0.90; gen --lines X --redo.
• Line too long for a bucket → plan_line errors past 10 s; split it. A 10-word line at 4 s gets cut off (10-06): budget 6 s.
• Real clip with burned-in captions or the wrong kid age → swap the clip; Kris has teen footage, Jessica/Lori younger boys.
• Provider refuses the face → another ref/route. Never blur, crop or distance a face to get past a likeness filter.
• ~1 in 10 avatars won't animate → new reference frame; don't fight the model.
• Fallback A, real creator + AI b-roll: cut a real Dawn creator's lines (clip library, or a new creator via ugc-engine/sops/04-outreach-sop.md: perpetual paid usage, $200/video cap) into aroll/<id>.mp4, set text to her real words, pick --name avatar a still from her clip, then verify → assemble. Confirm usage rights with Fish.
• Fallback B, animated (cartoon-h3 narrator): EL mom narrator over literal shots, no lip-sync, band from the cartoonified ref with the wrist rule; pack per format-spoken-drama §6 minus voices:. Weakest prior (short cartoon VO = 0 purchases on n10).
• NOT BUILT (needs Fish's go): a GMV-ranked TikTok Shop sourcing tool; in-model extend.

13. Sources

• davidaistar: ~/research/davidaistar/REPORT.md; analysis/gemini_Rqim7SlMtZg.md, gemini_gpzy2ofG0Dw.md, gemini_PmXqgiFn9_Y.md, gemini_CZwfTy7jZcQ.md, analysis/ugc_seedance.md, analysis/diff_ai_ugc.md, analysis/diff_cloning.md, foreplay/06_yapper.gemini.md; transcripts in txt/.
• Our code: skills/ugc-omni/SKILL.md, scripts/ugc_omni/omni.py, scripts/yapper/yapper.py, scripts/clip_library.py, scripts/register_dawn_broll.py, scripts/cartoon_h3/endcard.py.
• Dawn data: MAP.md, output/dawn-combo-map-2026-10-05/aggregate.txt, product-truth, claims-register.
• Memory files in §1, and project_ugc_engine.md.

Not built yet (needs Fish's go)

Every line in this skill that mentions something not built, roughly de-duplicated. Some items repeat in different words.

• NOT BUILT (needs Fish's go): a GMV-ranked TikTok Shop sourcing tool; in-model extend
• Motion transfer / recast is not wired (registry: NOT BUILT)
• Extend as a candidate role (registry: video extend, NOT BUILT)