Short ASMR / satisfying-sound contrast ad for Dawn (10-30 s): a deafening alarm cacophony (phone alarms, a bell clock, a smoke alarm, mom's knock) that wakes the whole house except one sleeping teen, then dead quiet and ONE soft buzz on his wrist that wakes him and nobody else, with close-up textures and crisp foley.

Lane: cartoon-h3 narrator (stylized video role) by default; photoreal object cut (cheap i2v role) as a test cell · not built yet: 4 (see below) · source: /Users/ayden/.openclaw/workspace/skills/format-asmr

ASMR — sources

Reference ad (paraphrased; structure only)

davidaistar short YtHf2oQisLQ "Claude creates ASMR ad with unbelievable physics" (~/research/davidaistar/txt/YtHf2oQisLQ.txt). It's a clip cut from the long video f83UH-qmNUs ("Claude Code + Seedance 2.0 = automated UGC", analysis/gemini_f83UH-qmNUs.md), shown there among the bonus generations. What the viewer sees: a soda poured into a glass packed with ice, close, with the ice rising as the liquid fills. No person, no words. The whole hook is believable physics plus pour/fizz sound. His claim is that a drink brand could "pump out hundreds" for organic reach. The model is unnamed in the short; the parent video ran Seedance 2 on StarPop, 9:16, 14 s, 720p. Prompts were written in blocks: Action / Audio (voice type, room tone, delivery, dialogue) / Quality (resolution, lighting, texture).
Keep: one physical action shot close; sound as the hook; texture you can almost feel (ribbed silicone, sheet ripple); the Action/Audio/Quality split as the motion-prompt checklist (every buzz prompt says what moves, what stays still and what it should sound like, even though we lay the sound ourselves).
Change: his is pure product porn with no problem and no person. Dawn's core bans mystery openers and character-less explainers, so we add a symptom header on frame 0, a sleeping teen, and her as the one the din wakes. His physics carries the ad. Ours is a contrast (loud fails, quiet works) that proves the mechanism. His is an organic-reach claim with no paid numbers, so treat it as untested for paid.
Shorts table (analysis/guide_copy_images_shorts.md): the row for YtHf2oQisLQ files ASMR as "a scalable organic-content format for beverage brands, on physics realism alone". Cross-video patterns used here: #4 improvise-first, refine-second (let the model improvise the macro, then fix only the band in an edit pass); #8 pick the model by task (multi-step physical demos → Seedance-class i2v, here the Cheap simple-motion i2v role for object macros); #10 Claude writes the structured prompt and the model improvises the rest.
Closest Dawn precedent: output/concept-packs/dawnbands-wakes-everyone-but-him-2026-08-17c.md ("It wakes everyone but him", claymation, 62 s, shipped 8-17 on the old seedance lane). It sold the household damage his alarms do to everyone else, with 11 per-scene SFX at doctrine levels via scripts/dawnbands_wakeseveryone_finalize_2026_08_17.py → scripts/finish_video.py --sfx-plan, and a near-silent scene on purpose ("the quiet is the beat"). Keep: the angle, the cue-per-scene plan, the levels, the deliberate quiet. Change: 20-25 s instead of 62; the cartoon-h3 lane (that finisher reads the old manifest.json/pipeline-state.json runs, not cartoon-h3 run dirs); its SFX kit dir no longer exists.

Research diffs folded in

Diff	Row	Where it lands
`analysis/diff_animation_styles.md`	per-style model routing; cheap path for simple motion	SKILL §9: photoreal object cut runs on the Cheap simple-motion i2v role, the cartoon cell on the stylized-video role
`analysis/diff_animation_styles.md`	"export without per-clip QA" = don't adopt	SKILL §10: by-eye pass on every buzz clip (Gemini ~1 fps can't see it)
`analysis/diff_ai_ugc.md`	omni-reference for product accuracy is unproven on our account (3-4/10 in yapper)	SKILL §7: object cut uses real product photos as start frames; no hand-held band renders
`analysis/diff_ai_ugc.md`	face-blur filter bypass = never adopt	photoreal cell is objects-only; no people, no workaround
`analysis/diff_ai_ugc.md`	cheap-tier-first culling; short 15-30 s formats are cheaper per test	the whole format: ≈$3-5 per ad, one batch for the short cell

Our code receipts (read 2026-10-06)

scripts/cartoon_h3/assemble.py: fit_clip runs ffmpeg -an on every clip (clip audio dropped); captions from lines.json; master_mix(rendered, final, None, bed_tension=music, hit=None, riser=None) = ONE bed. DEFAULT_MUSIC = gentle-piano-ambience; run.py --music swaps the file.
scripts/vo_audio_chain.py master_mix: the bed is -stream_loop -1, faded in over 0.8 s, gain-normalized to BED_LUFS = -26, sidechain-ducked under the VO; master loudnorm to -14 LUFS / TP -1. So a din passed as --music comes out quiet, faded and ducked. Hence the post-assemble overlay.
scripts/finish_video.py: the per-cue mix pattern (volume, adelay, amix … normalize=0, -c:v copy) the beat-sheet overlay copies; its SFX_KIT dir is gone.
audio-library/INDEX.md + ls audio-library/sfx: ambience ×3, tactile ×1 (gentle-tap 3.4 s), transitions ×6. No alarm, knock, smoke or buzz files.
ElevenLabs /v1/sound-generation is called inline in scripts/zakdfilms_pipeline.py and scripts/higgsfield_pipeline.py (old lanes, per-scene); there's no standalone CLI.
scripts/cartoon_h3/vo_check.py: narration floor 3.4 wps, dead air 0.40 s; pack front matter min_wps / dead_air override them (meant for packs that mirror a measured source ad). This skill keeps the defaults.
scripts/cartoon_h3/pack.py: role words reveal / cta set the product flags (reveal/CTA batches ≤3 panels); intents grammar N. **role** — text.
scripts/seedance_img2vid.py: --image --prompt --output --duration (4-12) --resolution (720p|1080p) --fal-key; endpoint fal-ai/bytedance/seedance/v1.5/pro/image-to-video; sends no audio parameter.
scripts/register_dawn_broll.py --dir + scripts/vo_broll_runner.py --pack: the existing real-footage-to-VO assembler; generated clips are untested through it.
knowledge/ad-formats/_MODEL-REGISTRY.md: roles used here are Stylized video (H3-max), Cheap simple-motion i2v (Seedance 1.5 Pro i2v), Sound effects / foley (proven = library files; candidates sonilo video-to-sound-effects, mirelo sfx v1.5 video-to-audio, kling video-to-audio, sonilo text-to-sound-effects, stable-audio-3 sfx; NOT BUILT in assemble), Video understanding (Gemini, ~1 fps).

Memory receipts

project_dawnbands_broll_seedance_playbook.md: vibration language under-renders; wake scenes must be reaction-primary; product-alone buzz renders as "shivers in place against the wood, rattling with tiny rapid movements"; fidelity block (diagonal ribbing, matte black inside and out, no smartwatch module); physics anchors; sleepwear stated on every in-bed human; watch-face drift (8-17c S9 re-render).
feedback_cartoon_vsl_sound_pacing_doctrine.md: SFX planned per scene up front; levels 0.20 / 0.11-0.13 / ≤0.08-0.10 / ≤0.12; muffled realistic sources; no tape-rewind transitions.
feedback_ai_avatar_broll_lane.md: photoreal PEOPLE gates (motion proof, locked character, line-match ≥8/10, eyes over scores). The object cut avoids people, so these don't trigger; if a person is ever added, they all apply.
corrections-hot 10-05: start from the simplest working recipe (one start frame, plain prompt) and add layers only when a test shows the simple one fails; 10-04: judge "looks HD" by eye.
Stake bank output/creative-strategy-cooper-2026-10-05/dawn-stakes.md: #17 "Everyone was up because of her alarm except her" [survey #9099]; #9 slept through the fire department [reddit 1gila3x]; #11 fifty trips every morning [survey #10232].
MAP.md §2A mechanism row (ramp+silent 1.74 vs silent-only 1.49) and §2B recipe B; §4 img-c1-dressed-asleep CPA curve ($38 → $68).

starpop articles (creator's own write-ups, read 2026-10-06; ~/research/davidaistar/starpop/txt/)

Applied: facebook-instagram-ad-sizes-and-safe-zones + how-to-preview-a-facebook-ad-before-publishing → §10 placement gate (the article names the UI regions but gives no percentages; the 14/35/6 % Reels box is Meta's published spec, cross-checked 10-06). are-ai-generated-ads-allowed-on-meta → §11 upload hygiene (Advantage+ enhancements off, live preview day 1). meta-video-ad-metrics-hook-rate-hold-rate-ctr-cpa, how-much-budget-to-test-meta-ad-creatives → §11 read/floor lines where this format needed them.
Read, nothing new for this format: ai-ad-formats-guide (format ↔ job table, shot-list order, references before video: already core §4), fully-ai-generated-ads-are-here-to-stay (strategy essay), how-to-calculate-break-even-roas-for-meta-ads (proposed for core §6, not format-specific), how-to-add-correct-captions-in-any-language + how-to-fix-wrong-capcut-auto-captions (script-as-source captions = our assemble rule already; his SRT → CapCut/Premiere import maps to Remotion per core §2a).
Not adopted: (1) "an AI actor must not voice a fictional first-person experience / fabricated testimonial" (formats guide, AI-ads article, organic-vs-paid) — conflicts with Fish's claims unlock (10-02) and our dramatized mom-narrator recipe; risk is Fish's call, recorded here only. (2) His 25-char vertical caption line cap — VideoAd chunks by its own MAX_PHRASE_CHARS 28 + two-line fit, Fish-set. (3) Organic-first discovery (ai-organic-social-vs-paid-ads) — Dawn's message and checkout path are proven, Testing CBO is our discovery lane. (4) His model picks (Kling O3 Pro, Sora 2, P-Video) — registry roles only; untested candidates.