Two-mom podcast-clip ad, 25-60 s. One mom tells her blamed-for-the-mornings story at the mic, the other reacts and asks, and the band lands as the answer in a product-gated cutaway.

Lane: (A) cartoon-h3 dialogue (staged, no lip-sync); (B) photoreal = ugc-omni (one Omni take per line, two speaker frames, gates G0-G4); H3 native voice is a side arm only · not built yet: 0 (see below) · source: /Users/ayden/.openclaw/workspace/skills/format-podcast

Podcast — beat sheet, banter rules, cut list

Runtime target is 25-60 s, which is about 90-200 spoken words including turn gaps. Baseline is 45 s / ~150 words. Percentages are of runtime. Beats are ranges of pack lines and never split a sentence.

#	Beat	%	Lines (45 s)	Speakers	Job	Line rules
1	Cold open	0-10	1-3	MOM → HOST	Drop the viewer mid-conversation on her stake	Line 1 ≤10 words and names her situation by an observable symptom (awareness gate). The stake comes from the stake bank or VOC. HOST reacts within 2 s. No backstory, no product.
2	The mornings	10-30	4-10	MOM, HOST interjects	Prove she IS the alarm	Concrete counts and times from VOC (knocks, alarms, minutes). Failed solutions stay generic (louder alarm, phone across the room, lights on): never brand names, never bed shakers. HOST interjects at least twice ("Every morning?", "How many alarms?").
3	The blame	30-45	11-15	MOM ↔ HOST	Raise the stake; it costs her	Who blamed her (husband, ex, school, MIL) and what it cost her (letter, mediation, a fight in front of the kid). HOST raises it one step, and the step is a question, not a lecture.
4	Root cause	45-60	16-19	MOM (quoting the explainer) + HOST	Why he doesn't wake	The credible person sits INSIDE her story (nurse, pediatrician, OT) and MOM quotes them. Fish's ruled root-cause line goes in verbatim. At most 2 lines. HOST voices the objection ("But he has five alarms.").
5	The answer	60-78	20-24	MOM or HOST	Mechanism + product	Touch gets through; silent to everyone else; keeps getting stronger until he turns it off (ruled line). Product named ONCE. On screen: the band cutaway (product-gated omni b-roll, real footage, or the cutout graphic).
6	Vindication	78-93	25-27	MOM → HOST	Payoff on HER	The accuser concedes, or the kid gets himself up, quoted by MOM. HOST gets a one-line reaction. MOM gets the last word.
7	CTA	93-100	28	HOST or MOM	Offer	One line: product name + 60-night guarantee. End card (optional, `--end-card`; off by default per Fish 10-06) 3 s.

Banter rules (the script is most of the realism)

Short alternating turns. Most turns are 3-8 words; none is over 12. One sentence per pack line.
No monologues. No speaker runs more than 2 lines in a row. HOST speaks at least every ~6 s, even if it's only "No." or "Wait."
Reactions are lines. Tag 1-3 word reactions as their own pack lines ("Stop.", "Every day?", "Oh no."), so they get their own cut.
Interruptions. Write the cut-in as the next speaker's line, and end the cut-off line on an unfinished clause ("And then his dad says—"). In B, the overlap can be faked after assembly by J-cutting the next speaker's audio 0.15-0.3 s early. Never ask one generation to talk over another.
Questions carry the viewer. HOST's questions are the ones a problem-aware mom has, and at least one is an objection the band answers.
Said-out-loud test on every line. Pronoun-first for MOM. Allow at most one filler ("honestly", "like") per 30 s, and only if it's in the VOC. Spell numbers as spoken ("six thirty").
HOST is never a salesperson. She doesn't pitch, she asks. If she knows the product, she found it the same way (her own kid). No "I'm not sponsored" style lines.
A line that only works as a close-up of the speaker gets re-staged in A (listener or cutaway) or kept in B.

B: one ugc-omni line per turn (photoreal)

Every pack line = one lines[] entry = one Omni take on its speaker's frame (frame: "avatar" = MOM, frame: "host" = HOST). lines[] order IS the cut order: omni.py assemble concatenates the trimmed takes, so the intercut builds itself.
Short reactions ("Stop.", "To you?") are their own lines. Omni's 4 s minimum is padded with throwaway words and verify trims at the last real word, so a 1-word reaction costs $0.315 and plays at its real length.
A turn longer than ~4 s gets a b-roll line (story cutaway, visual: "broll:<id>") on top; the speaker's audio runs underneath. A product line is always visual: "broll:band" (product-gated omni b-roll, or real footage).
Every hook line sets end_frame to its own frame (flash otherwise lands it on the body's first frame, which may be the other woman).
J-cut overlaps (next speaker 0.15-0.3 s early) are not in assemble; add them on the final cut with ffmpeg only if Fish wants them after G4. - Split-screen (optional): the Remotion PodcastTwoShot stacked layout on the per-line aroll/<id>.trim.mp4 takes (python3 scripts/remotion_stitch.py comp PodcastTwoShot <props.json> <out>).

Structure example (ILLUSTRATIVE: shows the shape only, not shipping copy; real lines come from the stake bank, VOC and Fish's approval)

[MOM] "The school mailed the truancy letter to me." (cold open, stake on her)
[HOST] "To you? He's the one asleep."
[MOM] "Fifteen tardies. My name on the envelope."
[HOST] "How many alarms does he have?"
[MOM] "Four. They wake his sister. Not him." …
[MOM] "Our nurse said his brain tunes the sound out." (root cause, quoted from inside her story; real copy uses Fish's ruled line verbatim)
[HOST] "So louder doesn't fix it." …
[MOM] "It buzzes his wrist. Nobody else hears it." (answer; broll:band cutaway: band on his wrist on the pillow, display off, gated) …
[MOM] "His dad did one week of mornings. Then he apologized." (vindication)
[HOST] "Okay, I'm ordering one tonight."