format-song — accused, tested, vindicated (sung)

This skill WRAPS the binding playbook knowledge/ad-formats/song-ad-playbook.md (cited below as PB §n). It does not restate it: it adds the beat sheet with word budgets, the missing storyboard checklist, the variant cells from the 10-06 songs diff, and the code truths a run trips on.

1. Read first

• knowledge/ad-formats/_FORMAT-CORE.md (doctrine, lane limits, gates, launch). Binding.
• knowledge/ad-formats/song-ad-playbook.md in full (PB §1 concept, §2 lyric red-team, §3 Suno, §4 art styles, §5 scene types, §6 production, §7 launch). Binding.
• references/beat-sheet.md (skeleton, % + word budgets, line rules, short-song cell), references/sources.md (reference ads, diff rows, receipts).
• output/dawn-combo-map-2026-10-05/MAP.md §2A (what wins) + §3 (the accuser queue) and the stake bank it points to.
• Song copy doctrine: ~/.hermes/skills/openclaw-imports/direct-response-copywriter/references/song-dr-ballad-scripting.md; Evolve rules per core §1.
• Voice: output/dawn-iterate-2026-10-04/video/song-voice-refs.md (style strings A/B + negatives). Last run: output/dawn-iterate-2026-10-04/video/RUN-LOG.md.
• Product: knowledge/brands/dawnbands/product-truth.md.

2. When to use / when not

• Use for every new Dawn song: iterate a paying song's LOCKED beat sheet and change ONE thing (PB §1), normally the accuser + the verbatim line, taken in MAP §3 order: school accuses her (truancy letter) → CPS / school social worker → her own kid ("scared of me in the mornings") → her boss.
• Evidence status: signal, not proof. Exhibit A (sa-ontime-10-exhibita-a) $33.7 CPA, ROAS 2.48 on $292; oneweek (sa-ontime-04-oneweek-a) ROAS 1.81 on $1,106; sl0919-h2 $40 CPA at 2:40. Every song number sits on <$1.2K spend. Songs: CPM $49-68, CTR 3.8-4.7 %, frequency builds ~half as fast per $ as statics (MAP §4).
• Don't use for callout/brand-voice angles (recipe B → statics/natives), a brand-voice or character-rap song (SLOP SONG GOAT 0.72), a song with no accuser on mom (the-shirt, drain-mom/dad: 0 purchases), kid-money/college stakes (0.40, 0.58), stakes the mornings didn't cause, or another dad-POV song before the existing one reads. To test song vs story, use format-spoken-drama.
• Husband/MIL songs already exist (five mornings, in my day, always nagging, dad-POV another week) and were never read: the cheapest next read is their own floored ad sets (Meta write, Fish's go), not a new build (MAP §3.5).

3. The format

• 9:16, 3:00-5:30 sung (10-04 promoted songs 336.7 / 228.5 / 262.8 s, finals after --song-end 329.3 / 226.6 / 250.8 s; exhibit A 271.5 s). One art style per song (disney3d, skeleton, claymation, pixar, paper), anchored on a real winner frame (PB §4).
• The viewer hears ONE continuous song: a mom 35-54 singing her own story, first person, talk-sung verses + plain belted chorus. Nobody on screen sings (Fish 9-19). The picture shows, literally, what each line says.
• Hook header on frame 0 (--title-style dragon, held for the first sentence ≥4 s), first sung word ≤0.3 s, captions = typed lyric lines on the sung timings, motion end card (optional, --end-card; off by default per Fish 10-06) over the last 3 s.
• Reference shapes: our exhibit A (structure LOCKED); Foreplay 12_musical (Resilia, 7:25 narrative VSL: mentor + mechanism + revenge) and 13_suno (Smooche, 5:37 Pixar musical: humiliation → discovery → the rival asks her secret).

4. Why it sells

• Recipe A (MAP §2A): someone close blames HER for the mornings → resentment/humiliation → a witness makes it public → she hands the mornings over and he fails → a credible person inside the story names the root cause → discovery → vindication in front of the accuser. The band is the evidence that clears her, not a pitch.
• Problem-aware: line 1 is the accuser's verbatim line or her stake, in action. The song carries ~4 min of attention at half the frequency build, so one concept lasts longer per dollar (theory, unproven past ~$550/week).

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

Budget a ~4:30 song at ~2.0-2.5 sung wps (Suno sings ~1.9-2.6 regardless of the prompt) = 550-630 sung words, the parent's count (PB §2).

6. Pack spec

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

• Front matter (core §5 + song keys; model on output/concept-packs/dawnbands-song-nottardy-2026-10-04.md): format: cartoon-h3, format_skill: format-song, brand: dawnbands, parent: (paying song + length/words + CPA), iteration: (the ONE variable), title: (header, | = line break), overlay:, end_card: Dawn Band. It wakes him, not you. 60 nights to try it., optional end_card_secs / end_card_clip / end_card_product, style + style_key (a run.py --style choice: claymation, pixar, disney3d, skeleton, paper, bergamot, bluey), style_block, visual_negative, style_anchor, audio: song (kie.ai Suno), suno_model: V5, suno_style, suno_negative_tags, suno_vocal_gender: f, cast, envs, product_truth. New slug per run.

• Lyrics: ## Suno lyrics prompt (custom mode) fenced block (song.py's exact regex) with [Verse N] / [Chorus] / [Bridge] / [Outro] tags only; no [Spoken] tags (PB §3).
• ## VO Script: N. "lyric line" one sung line per pack line, identical to the lyric block, 5-9 words, numbers spelled out. No [SPEAKER] tags in song packs: a song is one sung voice; tags would split it into dialogue lines. (The splitter's old "mouth moving" speaker rule was removed 10-06; tagged lines are now staged as voice-over.)
• ## Scene visual intents: N. **narrator** — <env>: <named people, doing what, framing>; product lines **narrator (product reveal)** or **cta** (batch ≤3 panels). Every person by name or sheet label; side characters "(not Mom)".
• <slug>.args beside the pack: --char-ref "Label=path", --env name=path, --product-ref, --characters "..." (read by packargs.py).
• Cast: Mom (avatar: ordinary American mom 35-54) + accuser + explainer + kid (+ witness). Check the cast against the avatar before sheets (PB §4: reused sheets carry their old look). Envs: 6-8 rooms (PB §5).

7. Visual grammar

• Shot types per beat: PB §5 table (accusation two-shot, evidence insert, wide wake routine, her cost, witness room, low-point close-up, reversal pings, authority two-shot, discovery, proof, vindication, quiet payoff). ~55 % wide / 35 % close / 10 % medium; 85-98 beats; ≤4 panels per batch.
• Song storyboard gate. Since 10-06 the splitter (storyboard.py _SYS) carries these rules and python3 -m scripts.cartoon_h3.check_plan --lint output/cartoon-h3/<slug> --pack output/concept-packs/<slug>.md checks every beat (mouthing/sings/mid-sentence, speech verbs, "narrator", lit/glowing band, smartwatch not in the intent, raised/looked-at/pressed wrist, product before the first reveal/cta beat, readable text/clock numerals, bedside wakes, beats <1.2 s; exit 1 on any BLOCK; unnamed people / non-wide wakes = WARN). Lint to 0 BLOCK, then read EVERY beat against the lyric for what a regex can't see, before storyboard.approved:
• [ ] No mouthing: no "mouth moving", "mid-sentence", "sings", "speaks", "says to camera". Yells/shouts as actions are fine.
• [ ] Every person named (Mom, Caleb, the attendance officer) so the hero sheet attaches; never "a woman", "a mother", "relatives", "the narrator", "blurred figures". Side characters carry "(not Mom)" / "(not Caleb)".
• [ ] Wake shots: WIDE from the doorway, teen fully dressed ON TOP of the duvet, adult at the light switch/blinds; never an adult at the bedside, never a teen under covers, no shower (closed door + steam).
• [ ] Wrist rule: wrists down or resting, band side-on, nobody looking at the wrist; "held up" = arm extended toward people at chest height; presses on the band happen with the band lying on a nightstand; product beats in daylight.
• [ ] No product before the discovery verse (first band beat = beat 8; no band on anyone's wrist earlier).
• [ ] No readable text: papers = yellow stripes, phones/laptops turned away (glow only), clocks hands-only, signs blank, no logos.
• [ ] No split screens / phone-call split frames; two consecutive near-identical shots get different framings (H3 stacks them).
• [ ] Similar-age characters in different batches or visually distinct (hair, colour, build).
• [ ] Buzz shown as eyes opening + cartoon alarm lines; never "vibrates / strong pulse / twitch / jiggles" next to a teen in bed (fal content-policy reject).
• [ ] Real-life plausibility per PB §2 (shift times, who can see what, a mediator doesn't "order" custody).
• [ ] Frame 0 states HER, not the room: beat 1's first panel = Mom + the stake object (the letter, the exhibit) in readable light, a thumbnail that names the persona; never a dark establishing bedroom. The dragon header sits over her face, not over an object or empty set (davidaistar's own mattress song opened on a dark bedroom: 2.05 % thumbstop, his #1 fix; starpop ai-singing-ad-mattress-case-study, how-to-edit-ai-song-ads-capcut).
• Edit beats in plan.json (keep a .bak), not the pack intents: editing pack visuals after the song is promoted breaks song_clock.

8. Audio

• Engine: Suno V5 on kie via song.py (PB §3). fal MiniMax Music / ElevenLabs Music are banned for this (wrong genre, word swaps, dropped lines).
• Style string rules: the voice is the avatar (Nashville country-pop storyteller woman 35-50, warm slightly husky chest voice, light twang, talk-sung verses, plain belted chorus, tail vibrato, no runs); each song gets its OWN string (genre/BPM/instrument differ: e.g. piano ballad 72 / acoustic pop-folk 96 / heartland country-rock 84); always "clear diction, every word easy to understand, vocals up front, dry clean mix"; ≤400 chars. Negatives always include R&B, trap, rap, breathy, whisper, heavy reverb, male vocal, choir, autotune, spoken word, mumbling, long intro (fuller list in song-voice-refs.md). No blank-line gaps or "beat switch" wording (Suno inserts 3-7 s pauses).
• Take check, then pick: song.py scores every take on arrival (output/cartoon-h3/<slug>/song/receipt.json: lyric_check PASS/REVIEW/FAIL, wps, first_word_s, missing/garbled/review lines). Then: listen at every flagged line (Scribe writes "6:30" for "six-thirty"; not an error); check short lines (≤4 words) word by word, since Suno flips "It was" → "It wasn't"; Gemini audio check (female lead, mom 35-54, clarity 1-10, sung not spoken); throw out male takes; measure gaps >1 s in take-N-words.json. Fix a misspelled brand/name in take-N-words.json BEFORE --pick.
• Promote: song.py --pick N --head 0.05 (trim keyed on the first LYRIC line, not the first ad-lib) → first sung word ≤0.3 s.
• End = the instrumental tail, kept: never hard-cut the resolution. Pick --song-end <s> at the musical resolution (Suno leaves 6-11 s of outro) so 3-5 s of instrumental follow the last sung word, then set --song-tail = pack end_card_secs = that tail (≥3.0): the motion end card (optional, --end-card; off by default per Fish 10-06) fills the whole resolving instrumental and the captions end before it (run.py warns if tail < card). Source: starpop how-to-make-ai-singing-ads-minimax-h3, how-to-edit-ai-song-ads-capcut (the tail is free attention at the moment the song resolves).
• No music bed, no SFX: the song is the full mix (run.py --song).

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

cd ~/.openclaw/workspace; P=output/concept-packs/<slug>.md; S=<slug>; T=output/dawn-iterate-2026-10-04/video
python3 -m scripts.cartoon_h3.song --pack $P --takes 2               # 2 kie tasks = 4 takes, ~12 cr/task (~24 cr) — Fish's go
python3 -m scripts.cartoon_h3.song --pack $P --pick N --head 0.05     # promote the chosen take (free)
python3 -m scripts.cartoon_h3.sheet <out.png> <prompt.txt> <style_anchor.png> [product-cutout-unlit.png]   # ~$0.22 each, Fish's go
A=(${(z)"$(python3 -m scripts.cartoon_h3.packargs $P --storyboard)"}); A=(${(Q)A})
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage storyboard        # free; run §7 checklist on every beat
touch output/cartoon-h3/$S/storyboard.approved                        # Fish only
A=(${(z)"$(python3 -m scripts.cartoon_h3.packargs $P)"}); A=(${(Q)A})   # all refs must exist now (core §5 packargs)
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage canvas --canvas-workers 4                 # paid, ~$0.22/batch
python3 -m scripts.cartoon_h3.run "${A[@]}" --stage gen --prompts-only                        # Fish reads 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                                     # paid, ~$2.5/song (fal upscale)
df -h ~ ; REMOTION_CHUNK_FRAMES=1500 python3 -m scripts.cartoon_h3.run "${A[@]}" --stage assemble --hd \
  --caption-mode tiktokbox --title-style dragon --song-end <s> --song-tail <tail = end_card_secs, 3-5>   # + (optional end card) + gemini_qa
python3 $T/tools/onset.py output/cartoon-h3/$S/$S-final-endcard.mp4 8                        # first vocal ≤0.3 s
mkdir -p output/cartoon-h3/$S/postqa; $T/tools/contact.sh output/cartoon-h3/$S/$S-final-endcard.mp4 output/cartoon-h3/$S/postqa/contact 2   # read every page

• Shot tiers (core §2b) + draft→final (core §4 gate 6b). Tag every batch A/B/C before approval: run.py does it itself (scripts/cartoon_h3/tiers.py, also python3 -m scripts.cartoon_h3.tiers <run>: A = first batch/hook, reveal/CTA/product beats, 3+ characters in a shot; C = no person in any shot; else B) into plan.json + <run>/tiers.md with the per-batch fal cost; re-tag in review if the word match is wrong. Per-tier models: run.py --tier-models "A=h3-max,B=h3-max,C=h3" (fal only; default = --model for every tier, i.e. all H3-max). h3 on B/C is a candidate until a one-batch side-by-side passes Fish's eye (registry "How to adopt"). Draft→final: run.py --draft = 480P ($0.05/s vs $0.08/s at 768P, fal page re-read 10-06) into <slug>-draft, seeded from the approved run (VO, plan, approval, canvases), never the shared-body cache, no upscale/end card (optional, --end-card; off by default per Fish 10-06); it prints the native 768P re-render command for the approved batches. Finals re-render natively, never a 480P→--hd upscale (Fish judged native sharper 10-04). A re-render is a new sample: judge staging and prompts on the draft, not the exact take.
• Never run songkie.sh / songfull.sh / songfix.sh / hookrun.sh for a Dawn song: they pass --yes, force --style disney3d, songkie.sh/songfull.sh auto-touch storyboard.approved, and songfull.sh/hookrun.sh hard-code the 9-19 cast.
• Lit band in a clip: pose fix + canvas re-render + --regen N first; free repaint after: python3 $T/tools/deglint.py <clip> <t0> <t1> <x0> <y0> <x1> <y1> [--hue=..] [--largest] (fails on skeleton bone). Sweep every product clip at 1 fps by eye.
• Cost per ~4-5 min song (PB §6): ~$31-43 on fal (H3 ~$22-30 at h3-max 768P $0.08/s + ~$0.10-0.16 reference surcharge per batch: nottardy's 24 batches quote $24.73 in tiers.md; canvases $6-10, upscale ~$2.5) + ~$8 sheets + ~24-72 kie credits. Ledger caveat: h3.py FAL_PRICE was fixed 10-06 from the fal page (it billed 768P at the 480P rate), so 10-04 ledger H3 numbers understate ~1.6×; run.py --dry-run quotes per batch. Regen budget 8 per song. Check kie/fal/Higgsfield balances before promising a render.

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

• [ ] Parent named with CPA + length/words; ONE variable stated; accuser line from raw VOC/stake bank (cited); mornings caused it; stake lands on her.
• [ ] Beat sheet filled within the word budget (550-630 long; 100-140 short cell); explainer is a person in the story; Fish's root-cause/mechanism lines verbatim.
• [ ] Logic pass → Opus red-team → fix → re-check the fixes (PB §2); Fish approves the lyrics.
• [ ] Own style string, distinct from the last 3 songs; negatives present; suno_vocal_gender: f.
• [ ] Takes: receipt lyric_check PASS (or REVIEW lines listened to), short lines exact, female mom voice, gaps <1 s, Fish listened to the pick.
• [ ] --pick N --head 0.05; first lyric ≤0.3 s.
• [ ] Cast checked against the avatar; sheets montage-reviewed (empty rooms, dark band, right outfits).
• [ ] §7 storyboard checklist passed on every beat; Fish touched storyboard.approved.
• [ ] Canvas montage eyeballed; motion prompts read (--prompts-only): no lit-digit, mouthing, "narrator" or "storyboard" wording.
• [ ] Smoke batch passed before the wave; balances checked; regen tally ≤8.
• [ ] 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; seam/split-screen check.
• [ ] Final: first vocal ≤0.3 s (onset.py), header on frame 0 over Mom (§7), captions = lyric lines, end card (optional, --end-card; off by default per Fish 10-06) over the whole 3-5 s instrumental tail with captions finished, contact sheet every 2 s read in full, gemini_qa scored (if the in-assemble call returns no score, run scripts/gemini_qa.py --asset <final> --type final standalone and read its flags).
• [ ] Caption safe zone: moved 10-06 (Fish approved "Move up to ≤62%"): tiktokbox and dragon blocks end at 64 % of the frame (centre ~62 %), tiktok centres at 59 %, all as a % of the height so 720 and 1080 renders match; the dragon header is moved from 8 % to 15 % at assemble (assemble.title_safe_zone). Finals assembled before 10-06 keep 68/74 % (re-assemble to move them; local, $0). Still view one final in Ads Manager's Reels placement preview before launch.

11. Variants & test plan (one variable per ad set; launch per PB §7 + core §6)

12. Failure modes

13. Sources

knowledge/ad-formats/song-ad-playbook.md; _FORMAT-CORE.md; ~/research/davidaistar/REPORT.md, analysis/diff_songs.md (rows 3-25, top 5), analysis/gemini_8NzJJ3e7uZ4.md, analysis/gemini_PmXqgiFn9_Y.md, foreplay/12_musical.gemini.md, foreplay/13_suno.gemini.md; output/dawn-combo-map-2026-10-05/MAP.md; output/dawn-iterate-2026-10-04/video/RUN-LOG.md + song-voice-refs.md; scripts/cartoon_h3/{song,run,storyboard,check_plan,endcard,reskin,hooksplice,h3}.py; skills/cartoon-h3/SKILL.md; memory project_song_lane_2026-09-19.md, reference_kie_ai_suno.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.

• 3 hooks via kie replace-section first verse only, same voice/key/tempo ≥10 s window; Suno re-renders ~30-60 s around it, swap the WHOLE song, regenerate only batches inside the window (~$3-5 + ~20 kie cr per hook) Not built yet (needs Fish'
• Sung chorus on camera 2-4 chorus beats become Mom singing to camera same audio + beats via reskin machinery + vocal stem Fish ruled against it 9-19. Test cell only with his explicit go; not built: no lipsync beat flag, h3.py tells H3 "no li
• 25 (P0) Splitter _SYS has none of the song rules; check_plan.py checks only timing/caps; ~100 beats hand-fixed, ~11 lit-band regens on 10-04 SKILL §7 manual checklist; §12 "not built" line
• 21 (P1) One hook per $26-33 song SKILL §11: hooksplice (built) + replace-section (not built)
• 15 (P1, Fish call) On-camera sung chorus untested; overrides 9-19 ruling SKILL §11 test cell, needs Fish's go, not built
• 14, 17, 10, 11 (P2) Song-aware H3 prompt, ≥0.95 slow-down, beat-grid snapping, "no extra in-between panel" canvas line not built; out of scope for a format skill (code changes, Fish's go)