format-warehouse — the pile is the proof, her morning is the story

1. Read first

• knowledge/ad-formats/_FORMAT-CORE.md (doctrine, lane limits, shot tiers §2b, gates, launch) and knowledge/ad-formats/_MODEL-REGISTRY.md (roles → models). Binding.
• references/beat-sheet.md (skeleton, word budgets, footage-truth table, shot tiers, grounding procedure), references/sources.md (evidence status, what was checked, receipts).
• Lane: skills/vo-broll/SKILL.md (EVOLVE: 3 hooks + 3 bridges, ONE shared hold + CTA, ≤45 s, no talking heads) and skills/vo-broll/references/production-handoff.md (clock + footage laws, runner gaps). EVOLVE structure is binding inside this lane; this skill adds the warehouse layer on top.
• Brand: knowledge/brands/dawnbands/product-truth.md (facts), knowledge/brands/dawnbands/claims-register.md (claims unlocked 10-02), output/dawn-combo-map-2026-10-05/MAP.md §2B (recipe B), copy rules knowledge/evolve/copywriting-rules.md.
• Memory: reference_dawnbands_packaging.md (POUCH, not a box), project_dawnbands_broll_seedance_playbook.md (b-roll prompt formula), project_vo_broll_lane.md (lane history), feedback_guarantee_positioning.md, feedback_ai_avatar_broll_lane.md (photoreal gates).

2. When to use / when not

• Use to test demand/social proof as the frame of a recipe-B video: brand voice straight to her, her daily burden, root cause + mechanism, social proof + guarantee as the authority, kid-independence payoff. Also when Fish wants a cheap, faceless, product-heavy video that needs no actor.
• Evidence status: UNGROUNDED. Fish's format #11 link was a duplicate of #10 (foreplay/links.txt has no 11; REPORT.md: "#10 and #11 are the same link"), and 10_character_arguments is the Penrose lotion-vs-cologne debate, not a warehouse ad. This framework is built from how the format commonly runs plus our own data. Ground it with §11's step the moment Fish sends the right link.
• What Dawn data says about the parts (MAP aggregates, aggregate.txt): social-proof authority in statics $17,469, ROAS 1.83 (4 W / 1 L) vs borrowed-person $27,569, 1.50 (5 W / 3 L); in songs social-proof is 1 ad, $182, 0.72. "Proof-led" angle = 1 ad ($468). Guarantee-led headlines lost (feedback_guarantee_positioning). HiStrips saturates "tested on 10,000+" claims, so a demand number alone fails the competitor-swap test.
• Don't use for recipe-A stories (accuser, vindication: format-song / format-spoken-drama), for an ad that is ONLY the warehouse (character-less explainer VO = 0 purchases, core §1 don't-build), for fake scarcity, or for any line that presents AI/stock footage as Dawn's real facility or staff.

3. The format

• 9:16, 30-45 s (vo-broll cap 45 s). Brand-voice narrator ("we") talks straight to her. Picture alternates two worlds: the packing room (pouches stacked on wire shelving, hands packing, a counted pile, the band buzzing on the bench) and her morning (the closed door, her third knock, the kid asleep, then the kid up on his own).
• Split rule: packing-room footage ≤50 % of runtime; her morning ≥40 %. The warehouse is the proof, never the story.
• Captions from the script on every word; a hook headline on frame 0 held ≤3 s; end card (optional, --end-card; off by default per Fish 10-06) ≤2 s under the last "tap below" phrase (handoff clock law 5).
• Three executions (pick one per ad; the choice is the test variable, §11):
• W1 faceless packing room (default): hands only, no faces, VO + b-roll on the vo-broll lane.
• W2 animated packing room: cartoon-h3 narrator; reads as illustration, so "this is our packing table" is fair storytelling.
• W3 on-camera packer/founder (test cell): a person says the lines. It's a borrowed person (statics 1.50). Production = ugc-omni (THE photoreal engine, Fish 10-06; gates G0-G4, §10); job shape references/omni-job.example.json. Unnamed, never a real employee's name, never billed as a real staffer; "Dramatization" label (Fish 10-06).
• Reference shape (UNGROUNDED): closest analysed ad is 05_founder (Azure Batana, 3:10: insider spokesperson + fast b-roll, cut back to the speaker every 3-4 s, mechanism, mother-proof, BOGO + guarantee). We keep the fast b-roll and the proof-then-offer close; we drop the talking spokesperson (W1) and the 3-minute length.

4. Why it sells

• Mechanism: volume = other parents already solved this, so she's not the only one and it isn't a gamble. A sophisticated market discounts promises but still responds to proof moments (feedback_guarantee_positioning). The pile is a proof moment she can see.
• Recipe B (MAP §2B): brand voice, her daily burden, exhaustion → relief, authority = social proof + guarantee + science (never a borrowed person in W1), payoff = kid independence or calm mornings.
• Why the warehouse alone isn't enough: "10,000 parents ordered it" is the leader's claim too. The swap-proof part is OUR root cause + mechanism lines over our footage (the band shivering silent on the bench while the room stays quiet). Concept formula still applies: proven structure + verbatim VOC line + her belief + the buyer who converts.
• Awareness: problem-aware. Lines 1-2 name HER by an observable symptom. "We can't keep these in stock" alone is a mystery opener and fails the core awareness gate.

5. Beat sheet (full rules, word budgets, shot list in references/beat-sheet.md)

Budget 30-45 s at ≥3.0 wps (lane floor) ≈ 100-150 words. Slots are ceilings; the voice sets the clock (handoff law 1).

6. Pack spec

• Step 0 — combo plan (feedback loop, workflow-combo-loop): python3 scripts/combo_loop/plan.py --format format-warehouse --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.

• W1 (vo-broll) = one JSON pack per hook variant (A/B/C) in one batch folder output/vo-broll/<batch>/, read by scripts/vo_broll_runner.py. Runner keys: ad_slug, pov (a VO_PROFILES key: default narrator-expressive; narrator-mom as a voice test cell), optional voice_id + voice_id_overrides_pov, scenes: [{scene_id, vo, shot: {source_pref: "rip"|"gen", search_terms, tags, gen_keywords}}]. Core §5 keys ride along as extra JSON keys the runner ignores but the gates read: format: vo-broll, format_skill: format-warehouse, brand, created, status, parent, iteration, title, overlay, end_card, execution: W1|W2|W3, footage_truth (per scene: real-dawn | stock | ai), stock_check (date + result of the live variant check).

• Scene ids: h1-hook, h1-bridge, hold-objection, hold-claim, hold-proof-reveal, hold-benefit, cta. The FIRST scene id containing reveal marks the reveal; the runner sinks product-tagged clips in unpinned scenes before it. Pin every cut anyway (<slug>-overrides.json: {scene_id: ["clip-id", "clip-id@sec", {"clip": "clip-id", "at": "word"}]}). Pins skip the pre-reveal filter, so never pin a pouch/band clip before hold-proof-reveal unless Fish has waived his 8-15 "no product b-roll before the reveal" rule for this ad (record it in the pack).
• Shared hold + CTA: identical scene ids, vo, pins and overlays across A/B/C. The runner has no shared-module object; copy A's built shared scene folders + state into B/C per the handoff, then build only hook + bridge.
• W2 (cartoon-h3 narrator): markdown pack per core §5 (format: cartoon-h3, format_skill: format-warehouse, style, style_anchor, cast, envs: packing-room, hallway, teen-room, kitchen, front-door, end_card); untagged narrator lines; build args with python3 -m scripts.cartoon_h3.packargs <pack.md>.
• W3 (ugc-omni job.json, skills/ugc-omni/SKILL.md owns the engine: read, never edit): style: photoreal, mode: cuts, aroll_model: flash, resolution: 1080p, image_arms: ["gpt2", "sd5"], product_refs: ["@dawn"], product_truth with ugc-omni's display-off sentence, core §5 keys + execution: W3, label: "Dramatization". lines[] = the W1 scenes in order: packer lines visual: "AI UGC" on avatar (packing bench, from a real warehouse-worker creator frame); her-morning scenes visual: "broll:<id>" (avatar: false); the reveal line product: true + end_frame: "holding" (band resting across his/her palm, close; product_action locked; the bench frame picked --product with the band lying on it); hooks carry end_frame = their own frame. Pouch shots stay real/composite (no pouch in @dawn refs).
• Cast (W1): one hand set for packing shots (same skin tone, same sleeve, no rings), one mom + one kid for her-morning cuts (handoff law 4: one actress, one kid per ad). No faces in the packing room.

7. Visual grammar

• Footage-truth table (full in beat-sheet.md): REAL Dawn packing footage → deictic lines allowed ("today's orders"). Stock or AI photoreal → the line carries the claim and the picture is illustration: no "this is our warehouse", no "meet Sarah from packing", no counted-pile-as-today's-orders, no overlay like "LIVE from our warehouse". Animated (W2) → deictic fine, it reads as illustration.
• Packaging = the stand-up ziplock pouch (blue-to-orange gradient, lowercase dawn wordmark, band photo, three bubbles). Never a rigid box, never a white box. Legible pouches = real footage or a composite of assets/dawnbands-pouch-3d-transparent.png (models garble its text). Generated pouches only out of focus, back side, or stacked at distance. Higgsfield element @dawn-packaging (44d7be04-db3a-4366-af74-7df20ea79d63) is the opt-in pouch ref on that backend.
• Band: plain matte black ribbed silicone, solid black inside and out, display DARK. Colours only if they're live variants (Black/Blue/Pink/Orange per the PDP JSON) and from real footage/refs. Never a smartwatch, never a lit face.
• Buzz: on the bench it renders as "shivers in place against the wood, rattling with tiny rapid movements" (playbook clip 29), the room silent around it. On a wrist: reaction-primary (arm jolt), wrist down or resting, band side-on (core §3 wrist rule).
• No readable text anywhere except the pouch: shipping labels, order screens, pick lists, whiteboards = blank or turned away (privacy + model garble). No branded outer mailers (we don't know what real orders ship in; ask Fish before showing one).
• Physics anchors on every product prompt: "lies flat on", "resting fully on", "closed fastened loop" (playbook). Hand-held band = closed loop flat in the palm, no face in frame.
• Shot tiers (core §2b): tier A = hands touching pouch/band, band on wrist, CTA, the hook's first 3 s; tier C = shelves, stacked pouches at distance, empty packing room, band on bench, clocks. Her-morning cuts come from footage first (own dawn-raw → stock), generated only when nothing fits.

8. Audio

• ElevenLabs via the runner profile (pov). Brand voice, warm and plain, not salesy; ≥3.0 wps on the final (handoff law 2; runner floor LANE_MIN_WPS 2.8), gaps ≤0.40 s on the voice stem. Script changes → rebuild with --force-vo (runner reuses VO without diffing text).
• One music bed, whole ad, ducked (runner master_mix, default audio-library/music-beds/warm-tactile/gentle-piano-ambience.mp3). Muted rip audio.
• The silent buzz is the point: no buzz SFX under the bench shot. The quiet IS the demo.
• W2: cartoon-h3 narrator rules (EL VO is the clock, vo_check ≥3.4 wps, end_card:).

9. Production (free until the first paid generation; every paid step is Fish's)

cd ~/.openclaw/workspace
curl -sL https://dawnbands.com/products/<live-pdp-handle>.js | python3 -c "import json,sys;d=json.load(sys.stdin);print(d['available'],[(v['title'],v['available'],v['price']/100) for v in d['variants']])"   # stock + price, read-only
python3 scripts/clip_library.py search --tags product,in-hand,product-pouch --keywords "pouch packing hands" -n 20   # own footage first
python3 scripts/register_dawn_broll.py --dir <folder Fish sends> --dry-run          # real packing footage, if Fish films it
python3 scripts/pull_broll_clips.py --packs output/vo-broll/<batch>/<slug>.json --out output/vo-broll/<batch> --sources pexels --dry-run   # stock for her-morning cuts
# generated b-roll, vo-broll lane = Higgsfield (Fish's standing 9-15 override in skills/vo-broll/SKILL.md: "do not fall back to fal.ai automatically")
python3 scripts/seedance_broll_runner.py --prompts <prompts.json> --run-id <id> --dry-run --no-discord               # cost estimate first (seedance1_5 = tier C)
python3 scripts/seedance_broll_runner.py --prompts <prompts.json> --run-id <id> --phase frames --no-discord         # start frames only → Fish picks (the draft)
python3 scripts/seedance_broll_runner.py --prompts <prompts.json> --run-id <id> --phase videos --no-discord         # paid, Fish's go; --allow-billing only if he says so
# fal route ONLY if Fish lifts the override for this batch: python3 scripts/seedance_img2vid.py --image <still.png> --prompt "<playbook formula>" --duration 5 --resolution 720p --output <clip.mp4>
python3 scripts/vo_broll_runner.py --pack output/vo-broll/<batch>/<slug>-A.json --no-stitch   # VO + scene build; read pace + gaps
python3 scripts/vo_broll_runner.py --pack output/vo-broll/<batch>/<slug>-A.json             # stitch + master mix
python3 scripts/gemini_qa.py --asset <final.mp4> --type final --style "photoreal b-roll"      # report the score, never adjudicate
python3 scripts/meta_preflight.py <final.mp4>                                                 # before any upload

• Model roles: tier C clips → Cheap simple-motion i2v (proven: Seedance 1.5 Pro i2v; in this lane on Higgsfield seedance1_5 via seedance_broll_runner.py, on fal via seedance_img2vid.py only if Fish lifts the override; candidates in the registry are untested: one-unit proof first, ≤~$2 without asking). Tier A band/hand clips → real footage first; else ugc-omni b-roll through its product gate (broll --stage frame → pick --name broll_<id> --product ≥7/10 + display state → --stage anim; display-off wording; worn or resting across a palm, never pinched upright). Seedance 2.5 product shots stay rejected (Fish 10-06). Pouch shots: real footage or the pouch composite.
• Provider conflict, flag once: skills/vo-broll/SKILL.md (9-15) carries Fish's explicit, standing Higgsfield override for generated clips; the registry (10-06) names fal as the general default but did not revoke it. The override wins until Fish says otherwise; ask at gate 1.
• Draft → final (core gate 6b): tier C = frames phase first (stills are the cheap draft, Fish picks), then videos; a 720p fal take re-rendered at 1080p is a NEW sample, not the approved motion. Generated tier-A band shots go through ugc-omni's gated exact first frame instead (no draft step: the frame check is the proof). W2: free drafts = storyboard, canvases, --prompts-only; the paid H3 draft is run.py --draft (480P), finals re-render natively (480P + --hd unproven, core 6b).
• W3 (ugc-omni, every paid step dry-runs until --spend): O=scripts/ugc_omni/omni.py; J=output/ugc_omni/<slug> → G0 python3 $O board --job $J → G1 refscan --src <real warehouse-worker video>, ref --pick, avatar --n 3 --arms gpt2,sd5, pick --name avatar, then variant --name avatar_band --product --arms gpt2,sd5 --prompt "<same person at the packing bench, the band lying flat on the bench. Display off: ...>" → pick --name avatar --product --file <that roll> (ugc-omni's continuous reveal: the band is in shot from line 1); variant --name holding --product ... → pick --name holding --product → G2 lock --line H1-1 --n 3 then the reveal line (product_action) --n 3, --approve each → G3 gen --lines all, verify, broll --id <id> --stage frame|anim (+ pick --name broll_<id> [--product]) → G4 assemble --hook H1, cost, scripts/meta_preflight.py. Cost: Omni $0.315 / $0.42 / $0.525 / $0.63 per 4 / 6 / 8 / 10 s, images ~$0.05-0.075; a 40 s W3 ad ≈ $10-13 with 2 lock rounds and regens.
• W2: the cartoon-h3 sequence in format-spoken-drama §9 (vo → storyboard → storyboard.approved → canvas → --prompts-only → --spend → assemble), narrator pack, no voices.
• Cost per W1 ad: $0 if own + stock footage covers it; ~6-10 generated tier-C clips at fal Seedance 1.5 Pro i2v pricing (read the fal page before quoting) + EL quota. W2 ≈ 4-6 H3 batches (h3.py FAL_PRICE h3-max 768P $0.75 per 15 s; the cartoon-h3 SKILL says $1.20 and core says the table may be inverted: read the fal page; run.py --dry-run prints the estimate). Check fal/Higgsfield/kie balances before promising a render.

10. Gates & QA checklist (on top of core §4 and the handoff's final QA)

• [ ] Parent named with CPA (the recipe-B static whose callout + mechanism the hold reuses); ONE variable stated; execution W1/W2/W3 recorded.
• [ ] Lines 1-2 name HER by a symptom; stake on her, caused by the mornings; VOC line cited; competitor-swap test failed (HiStrips couldn't run it unchanged).
• [ ] Root-cause + mechanism lines verbatim; demand claim as written ("10,000+ parents"), no invented order counts, per-day numbers or staff counts.
• [ ] Stock truth: live JSON checked today. All variants available → no sold-out / almost-gone / last-batch / restock lines. A restock or waitlist line only with Fish's confirmation of a real event, dated in the pack.
• [ ] Footage truth: every scene tagged real-dawn|stock|ai; no deictic place/people/volume line over stock or AI; no real names; no "live" overlays.
• [ ] Pouch, not box, in every packaging shot; legible pouch = real or composite; band dark, black, physics-anchored.
• [ ] Packing room ≤50 %, her morning ≥40 % of runtime (measured from the cut list).
• [ ] Price + guarantee match the live PDP; guarantee in the CTA, not the hook.
• [ ] One-unit motion proof on any NEW model before a batch (registry); W3: ugc-omni G0-G4, every gate Fish's (G0 script + board → G1 real ref + avatar + gated product frames → G2 the same line good 3×, then the reveal line with product_action → G3 every line + b-roll + band checked by eye → G4 the cut).
• [ ] Storyboard per phrase → Fish; generation prompts shown before paid video (feedback_motion_prompt_approval).
• [ ] No pouch/band shot before hold-proof-reveal (Fish 8-15) unless a waiver is recorded in the pack; generated tier-C clips go frames-first (draft) on the Higgsfield route unless Fish lifted the vo-broll override.
• [ ] Placement preview (starpop facebook-instagram-ad-sizes-and-safe-zones, how-to-preview-a-facebook-ad-before-publishing): scrub to the hook frame, the reveal frame and the end card (optional, --end-card; off by default per Fish 10-06) on a 9:16 overlay; header, captions, product and offer sit inside Meta Reels' safe box (top 14 %, bottom 35 %, sides 6 % are UI = keep content in y 14-65 %) and survive the Feed 4:5 centre crop (~15 % lost top and bottom). Remotion tiktok/tiktokbox/dragon captions centre at ~68-74 % = inside the Reels bottom band on Meta: flag it to Fish, never re-burn on your own.
• [ ] Final: 30-45 s measured, hooks 3-7 words inside 3 s, wps ≥3.0, gaps ≤0.40 s, freezedetect 0, 1 fps contact sheet eyeballed (label text, box, lit band, faces in the packing room), shared hold/CTA identical across A/B/C, Fish watched it.
• [ ] Ad link = Shop Now = any URL in text, one page (feedback_meta_link_match; scripts/meta_link_audit.py after launch).

11. Variants & test plan

12. Failure modes

13. Sources

~/research/davidaistar/REPORT.md (#10/#11 duplicate), foreplay/links.txt, foreplay/10_character_arguments.gemini.md, foreplay/05_founder.gemini.md, analysis/diff_ai_ugc.md, analysis/diff_cloning.md; output/dawn-combo-map-2026-10-05/MAP.md + aggregate.txt; knowledge/brands/dawnbands/{product-truth,claims-register,proof-bank}.md; skills/vo-broll/SKILL.md, skills/vo-broll/references/production-handoff.md; scripts/vo_broll_runner.py, scripts/ad_dissector.py, scripts/seedance_img2vid.py, scripts/seedance_broll_runner.py, scripts/clip_library.py, scripts/cartoon_h3/canvas.py; memory reference_dawnbands_packaging.md, project_dawnbands_broll_seedance_playbook.md, feedback_guarantee_positioning.md, project_vo_broll_lane.md, feedback_ai_avatar_broll_lane.md, reference_dawn_main_pdps.md. Details in references/sources.md.

NOT BUILT (needs Fish's go): a grounded reference ad (link pending); real Dawn packing footage (none in clip-library); hook headline overlay + end card (optional, --end-card; off by default per Fish 10-06) in vo_broll_runner.py (the 9-15 per-batch renderers aren't in output/ any more); a shared hold/CTA module in the runner; a 9:16 still CLI on fal; per-tier model routing (core §2b); an automated stock-truth guard (manual curl today).

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 grounded reference ad (link pending); real Dawn packing footage (none in clip-library); hook headline overlay + end card (optional, --end-card; off by default per Fish 10-06) in vo_broll_runner.py (the 9-15 pe