format-mini-movie — accused, tested, fooled, vindicated (chaptered)

format-spoken-drama is one arc in 2-4 min. This skill is that arc stretched into 6-9 chapters, each opening on a re-hook, with a subplot and a false solution in the middle. It is NOT a sung format (that's format-song) and NOT photoreal (§2).

1. Read first

• knowledge/ad-formats/_FORMAT-CORE.md (doctrine, lane limits, gates, launch) and knowledge/ad-formats/_MODEL-REGISTRY.md (roles → models). Binding.
• references/beat-sheet.md (chapter skeleton, % + word budgets, re-hook rules, illustrative structure), references/sources.md (reference ads paraphrased, what we keep/change, receipts).
• skills/format-spoken-drama/SKILL.md: dialogue staging, voices and pace keys are inherited from it unchanged; this skill only adds what length changes.
• output/dawn-combo-map-2026-10-05/MAP.md §2A (recipe A), §3 (accuser queue), §4 (frequency theory); stake bank output/creative-strategy-cooper-2026-10-05/dawn-stakes.md.
• skills/cartoon-h3/SKILL.md (shared body, hooks × leads matrix, credit discipline, VO clarity).
• Product: knowledge/brands/dawnbands/product-truth.md; claims: knowledge/brands/dawnbands/claims-register.md.

2. When to use / when not

• Use when a 3-4 min parent (song or spoken drama) already pays and Fish wants to know whether MORE story (chapters, subplot, a false solution) buys cheaper reach at a lower frequency build. One movie = one concept = one ad set.
• Evidence status: untested on Dawn. Signal only: Foreplay shows long AI narratives scaling for other brands (Rosabella 11:51 photoreal, Resilia 7:25, Smooche 5:37 sung). Our longest Dawn finals are the 10-04 songs (226-329 s). REPORT ranks this format "low until spoken drama (#4) reads".
• Build gate: do not start a mini-movie until format-spoken-drama (or a song at ≥3 min) has a 7-day read, so the test in §11 has a parent in the same audio form.
• Photoreal is out of scope. Gemini rated Rosabella 9.5/10 difficulty ("the absolute ceiling") for character/lip-sync consistency over 11 min. Our avatar-lane gates (core §3, feedback_ai_avatar_broll_lane) require a motion proof, ONE locked character at pairwise identity ≥8/10 on ALL shots, and a line-match audit before any i2v. A mini-movie has 4-6 people across 40-70 shots; the 9-30 Side Effects photoreal ad failed those gates with one person and ~550 Higgsfield cr. Re-open only after a photoreal format passes the gates and pays, via the registry's "Photoreal / multi-ref video with locked character" role on ONE 30 s chapter.
• Don't use for callout angles (recipe B), stakes the mornings didn't cause, kid-money/college stakes, a dad/kid narrator, or a character-less explainer.

3. The format

• 9:16, 5:00-12:00 (default first cut 7:00-9:00 ≈ 1,350-1,750 spoken words), animated in one style (claymation / pixar / disney3d / paper / skeleton), anchored on a frame from our paying song in that style.
• Chapters: cold open + 6-9 chapters of 40-90 s + close. No title cards (no readable text in frames; no mid-video overlay exists). The chapter is marked by sound and picture: a hard cut to a new room or day, and a re-hook line.
• Voices: Mom narrates in first person (~55-65 % of words); 3-5 other voices carry their key lines (accuser, witness/authority, explainer, kid, subplot person). Dialogue is staged un-mouthed (core §3).
• Captions from the script on every word; hook header on frame 0; motion end card (optional, --end-card; off by default per Fish 10-06) (end_card:).
• References: Foreplay 14_minimovie (Rosabella: a stranger's tragedy → the hero's own family at risk → rust/apple mechanism → reunion → doctor proof); 12_musical (Resilia: humiliation → mentor → mechanism → week-by-week log → the accuser sees it); 13_suno (Smooche: rival encounter → flashback → explainer met abroad → the rival asks her secret). Paraphrased in sources.md.

4. Why it sells

• Recipe A, extended: someone close blames HER for the mornings → it goes public and costs her → he takes the mornings and fails → a false solution works, then fails at the worst moment (the stake peaks) → a credible person INSIDE the story names the root cause → discovery → proof over days (subplot resolves) → vindication in front of the accuser. The band is the evidence that clears her.
• What length adds: the false solution is the objection handled in the story ("we tried a louder alarm / a smartwatch"); the subplot widens who the mornings hurt (her job, her kid's trust) while still landing on her; the long payoff gives the vindication room. Problem-aware from line 1.
• Media theory (MAP §4, unproven past ~$550/week): songs buy ~half the impressions per dollar and build frequency ~half as fast (lifetime 1.12-1.25). A 10-min film should sit at the far end of that curve; that is what §11 measures.

5. Beat sheet (full rules, word budgets and re-hook types in references/beat-sheet.md)

6. Pack spec

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

• One pack, one run (built path): the whole film is one ## VO Script, one ElevenLabs clock, one storyboard, one assemble. Chapter-per-run + concatenation is BUILT 10-06 (Builder C), chapter runs still need Fish's go: python3 -m scripts.cartoon_h3.concat --out <film.mp4> <ch1>-final.mp4 <ch2>-final.mp4 … [--xfade-frames 6] joins finished finals in one re-encode (size/fps matched to chapter 1, each chapter gained to -14 LUFS, burned captions untouched, seam times in <film>.concat.json). What stays true: each chapter keeps its own bed (it restarts at every seam) and its own QA, so one-pack/one-run remains the default. A ChapterCard Remotion composition exists (10-06: python3 scripts/remotion_stitch.py comp ChapterCard <props.json> <card.mp4>, then concat it between chapters) for a Fish-approved title-card test cell only; the no-title-cards rule above is unchanged. A cut-down (e.g. 5-min edit reusing the 9-min clips) is also NOT BUILT: clip caches key on segment text, so a new pack regenerates its batches (EL per-line takes are reused).

• Front matter (core §5 + dialogue keys, model on output/concept-packs/dawnbands-s2-attendance-office-h1-2026-10-01.md): format: cartoon-h3, format_skill: format-mini-movie, brand: dawnbands, parent: (paying 3-4 min ad + CPA), iteration: (the ONE variable: "3-4 min single arc → chaptered 8 min, same accuser"), title, overlay, end_card, style, style_key, style_block, visual_negative, style_anchor, cast, envs, product_truth, voices: MOM=<id>, <ACCUSER>=<id>, <EXPLAINER>=<id>, KID=<id>[, <SUBPLOT>=<id>], vo_tempo: 1.08, gap_same: 0.15, gap_turn: 0.3, max_gap: 0.22, seam_gap: 0.25, optional style_<SPK> 0.2-0.35, body_start: (first line after the cold open, for 3 hooks). Copy product_truth from output/concept-packs/dawnbands-song-nottardy-2026-10-04.md (display OFF and dark in every shot), never from the 10-01 attendance-office packs ("top face that can glow").
• chapters: = documentation only (no code reads it), e.g. chapters: 0=1-4, 1=5-22, 2=23-48, ... so the storyboard review can be split by chapter.
• ## VO Script: N. [SPK] "line", every line tagged (narrator = [MOM]), 5-12 words, one sentence. Chapter headings as ### Ch3 — Then you do it lines are safe inside the section (the parser only ends a section on ##, only reads numbered lines).
• ## Scene visual intents: N. **role** — <env>: <named people on screen, doing what, framing>. Roles: narrator, dialogue, narrator (product reveal) (first band line), cta (close and any mid-roll CTA; batch ≤3 panels). Every non-MOM line names its staging: "cutaway: Mom's face while Dad says it, Dad's back to camera" (see §7).
• <slug>.args beside the pack: --char-ref "Label=path" (label = the exact name used in intents), --env name=path, --product-ref, --characters "...".
• Cast: Mom + accuser + explainer + kid (+ witness, + subplot person) ≤6. Envs 8-12 reusable rooms; a new room = a new sheet, so reuse rooms across chapters.

7. Visual grammar

• No lip-sync (core §3). Speech is shown on the listener, the speaker's back/profile in a wide, over-the-shoulder, a phone/door, or an insert. The splitter's old "mouth moving" speaker rule was replaced 10-06 by voice-over staging (storyboard.py _SYS). Still write every non-MOM intent as a cutaway (--no-llm uses intents verbatim), then python3 -m scripts.cartoon_h3.check_plan --lint output/cartoon-h3/<slug> --pack output/concept-packs/<slug>.md = 0 BLOCK before approval (same gate as format-spoken-drama / format-courtroom).
• Chapter seams: a hard cut to a new room, day or time (clock hands, light through blinds). A batch may straddle a seam (the seam becomes a named Transition); the re-hook line must be its own pack line (≥1.2 s) so it gets its own panel. Forcing a fresh batch at every chapter (chapter_break) is NOT BUILT; only lead_start/body_start/body_end force breaks.
• Consistency across 40-70 batches:
• One character sheet per person (python3 -m scripts.cartoon_h3.sheet), one style anchor on every canvas (style_anchor → --style-ref via packargs). Sheets attach by whole-word match of the label to the panel text (aliases only for son/daughter/mom/dad), so name every person the same way in every intent; side characters "(not Mom)".
• Ref budget: with a style anchor, refs_for_batch keeps only 5 refs + the anchor (refs[:5]), in order characters → product → rooms. Plan each batch for ≤3 named people + 1 room (product batches ≤2 people + band + 1 room) or the room sheet is silently dropped.
• One outfit per person for the whole film (sheets lock outfits); no time jumps that age anyone.
• Two similar-age people (Mom + sister, two teens) contaminate each other: different hair colour/length, build and signature colour on the sheets, never alone together in a close shot.
• Batch seams are hard cuts by default; run.py --chain-frames (opt-in, 10-06) hands the previous clip's last frame to a same-scene/env seam as a start-frame ref. Unproven: one seam-pair proof on Fish's eye before a chaptered cut relies on it.
• False solution on screen: call it "smartwatch" or "alarm clock", never "band"/"wristband" (PRODUCT_IN_FRAME_RE would mark the batch product-visible and attach our band). In batches that show both, run.py adds the "bulky square-cased smartwatch, clearly not this band" clause.
• Product only from Chapter 6. Wrists down/resting, band side-on, display dark, nobody looking at it, product beats in daylight; buzz = eyes opening + cartoon motion lines. Wake shots wide from the doorway, teen dressed on top of the duvet. No readable text: letters are shapes with a red stamp, phones turned away, clocks hands-only.
• Framing mix: accusation/vindication on Mom's face; tests and proof days wide; explainer two-shot.

8. Audio

• ElevenLabs per speaker via voices:; Mom = the avatar's speaking voice (warm, slightly husky, 35-54); voices clearly different in age/pitch/energy. Explainer calm; accuser dismissive, never a villain.
• vo_check (automatic, before storyboard): an all-tagged pack is checked on the DIALOGUE profile only: overall ≥2.9 wps, any 6+-word line ≤4.8 wps, turn gaps 0.30-0.55 s, same-speaker 0.12-0.30 s, dead air ≤0.6 s, ASR round trip. Narration-heavy chapters should still read ~3.4+ by ear; nothing gates that per chapter. Never fix pace by speed; split or retake the line. Fish listens to the whole VO on the hub, by chapter.
• One ducked music bed for the whole film (assemble default); a chapter seam is not a music swap. SFX only if they don't mask words (no SFX path is built).
• End card (optional, --end-card; off by default per Fish 10-06) 3 s (end_card:); assemble ends 0.5 s after the last word and the card overlays the last 3 s, so the spoken close plays under it (write it as the final ≥2.5 s).

9. Production (free until sheets; every paid step is Fish's)

cd ~/.openclaw/workspace; P=output/concept-packs/<slug>.md; S=<slug>
python3 -m scripts.cartoon_h3.run --pack $P --dry-run                    # free before the VO exists: batch + cost estimate at 3.6 wps
A=(${(z)"$(python3 -m scripts.cartoon_h3.packargs $P --storyboard)"}); A=(${(Q)A})
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage vo                   # EL per line (quota) + vo_check gate
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage storyboard           # free; Fish reviews storyboard.md chapter by chapter
python3 -m scripts.cartoon_h3.sheet <out.png> <prompt.txt> <style_anchor.png> [refs...]   # ~$0.22 each, Fish's go
touch output/cartoon-h3/$S/storyboard.approved                          # Fish only
A=(${(z)"$(python3 -m scripts.cartoon_h3.packargs $P)"}); A=(${(Q)A})    # every ref must exist now
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage canvas --canvas-workers 4
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage gen --prompts-only   # Fish reads the motion prompts
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage gen --spend --backend fal --model h3-max --batches 1   # smoke
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage gen --spend --backend fal --model h3-max --gen-workers 8 --auto-regen 1
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage upscale --hd
df -h ~; REMOTION_CHUNK_FRAMES=1500 python3 -m scripts.cartoon_h3.run "${A[@]}" --stage assemble --hd --caption-mode tiktokbox --title-style dragon
output/dawn-iterate-2026-10-04/video/tools/contact.sh output/cartoon-h3/$S/$S-final-endcard.mp4 output/cartoon-h3/$S/contact 2

• Cost (scaled from the 10-04 ledger: 1,105 billed H3 s for ~810-830 s of finals ≈ 1.35× incl. regens; ~1.4 canvas renders/batch at ~$0.22; upscale ~$0.009/s; 10-04 songs ran 8.4-10.6 s per batch). H3 at the h3.py table rate ($0.05/s) vs the likely real 768P rate ($0.08/s, registry caveat; read fal's page before quoting):

10. Gates & QA checklist (on top of core §4)

• [ ] Build gate met: a same-audio-form parent has a 7-day read; parent named with CPA; ONE variable stated.
• [ ] Every stake and the accuser line sourced (stake bank / VOC, cited); mornings caused them; they land on her; subplot is mornings-caused too.
• [ ] Beat sheet filled within the word budget; every chapter opens on a re-hook and closes on a loop; mid-point false solution is generic (no brands, no bed shakers).
• [ ] Explainer inside the story; Fish's root-cause + mechanism lines verbatim; red-team done (timeline over days/weeks, who knows what, real-life plausibility, product facts).
• [ ] vo_check PASS; Fish listened to the full VO.
• [ ] Cast ≤6, distinct sheets, labels = intent names; style anchor present; per-batch ref budget planned (≤3 people + 1 room).
• [ ] storyboard.md: no "mouth", no "narrator", every person named, no band before Ch6, wrist rule, no readable text, no near-identical consecutive shots; Fish touched storyboard.approved.
• [ ] Motion prompts read; smoke batch passed; balances + disk checked; regen tally within budget.
• [ ] Every batch tagged tier A/B/C in storyboard.md; any draft→final pass followed §9's route (drafts never left in a shared-body cache).
• [ ] 1 fps lit-band sweep of every product clip; contact sheet every 2 s read in full, chapter by chapter.
• [ ] 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: freezes 0, dark 0, captions = script, header frame 0, end card (optional, --end-card; off by default per Fish 10-06) last 3 s, gemini_qa scored (standalone scripts/gemini_qa.py --asset <final> --type final if the in-assemble call returns no score), Fish watched the whole film. The Discord copy will exceed 8 MB at this length: share the hub/link.

11. Variants & test plan (one variable per ad set; core §6)

12. Failure modes

13. Sources

~/research/davidaistar/REPORT.md (§3 row 14), foreplay/14_minimovie.gemini.md, foreplay/12_musical.gemini.md, foreplay/13_suno.gemini.md, foreplay/04_drama.gemini.md, analysis/diff_songs.md (rows 13, 21-23), analysis/diff_animation_styles.md; output/dawn-combo-map-2026-10-05/MAP.md; output/dawn-iterate-2026-10-04/video/RUN-LOG.md §4 + plan.json stats; scripts/cartoon_h3/{run,vo,storyboard,assemble,pack,hooksplice,packargs,upscale,h3}.py; skills/cartoon-h3/SKILL.md; skills format-spoken-drama + format-song; memory feedback_cartoon_h3_shared_body_text_hash.md, feedback_higgsfield_credit_discipline.md, feedback_ai_avatar_broll_lane.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.

• One pack, one run (built path): the whole film is one VO Script, one ElevenLabs clock, one storyboard, one assemble. Chapter-per-run + concatenation is BUILT 10-06 (Builder C), chapter runs still need Fish's go: python3 -m scripts.cartoon_h
• Chapter seams: a hard cut to a new room, day or time (clock hands, light through blinds). A batch may straddle a seam (the seam becomes a named Transition); the re-hook line must be its own pack line (≥1.2 s) so it gets its own panel. Forci
• 3 Mid-roll CTA a spoken CTA at the end of Ch6 one cta line (product + 60 nights, PDP-true only) built (spoken). On-screen offer card mid-roll NOT BUILT
• 4 Re-skin art style new slug, same VO lines + intents, only the style keys changed: EL reuses the per-line takes (cached by voice + settings + text), storyboard re-splits, Fish re-approves manual route works; reskin.py is song-only (needs a
• 5 On-camera dialogue Lip-sync role (registry: sync-3 candidate) on 2-4 key lines one-unit proof first NOT BUILT, needs Fish's go
• 13_suno — Smooche colour-changing foundation, 5:37, 9:16, Pixar-style musical. Paraphrase: a 63-year-old woman runs into her ex and his much younger partner in a store → flashback to the divorce and makeup that failed her → a work trip, a s
• analysis/diff_animation_styles.md P1: last-frame chaining across batches (kills hard seams in continuous styles) is not built; at 60 batches every seam is a hard cut, so chapters lean into cuts
• Video extend: candidate for smoothing seams; NOT BUILT
• Lip-sync onto an animated frame (sync-3 candidate): the route to on-camera dialogue (cell 5); NOT BUILT