Podcast Ad

Podcast Ad

The ad is two people, one room, one conversation. Viewer overhears; nobody pitches the lens. Script is ~80% of the result.

Palmier does not generate a full multi-speaker room in one pass — segment the exchange into model-length beats, lock room/identity with matching reverse-shot stills, then assemble on the timeline.

Pair with video-prompting. For faces that must match brand talent across variants, run character-pipeline first. Finish captions/layout with ugc-editing patterns when helpful.

Why this format works

• She never addresses camera — kills “ad argument in the head”
• His doubts are the buyer’s doubts — conversion when he clicks
• Pattern-matches podcast clips in the feed, not classic ad shapes

Break overheard staging (lens eye contact, hard CTA, cheerleading sceptic) and you lose the format.

Step 1 — Write the exchange (sceptic first)

Sceptic’s four beats (required)

1. Open doubt — names the quiet part out loud
2. Honest question — hands her the floor without agreeing
3. The click — mechanism connects to his life (not a parrot of her claim)
4. Concession — stops arguing; does not gush

Believer craft: human comparison over feature list; disarm the scary version, then educate with a specific mechanism; callback close; no hard CTA inside the dialogue.

Map the script onto generation beats that fit list_models durations (often 4–10s). Prefer tight durations (often 5–6s). One conversational turn (or half-turn) per clip is safer than cramming. Flux 3 minimum is 5s.

Backchannels (reactive ad-libs)

Agent judgment (this skill): on longer believer turns, decide whether 1–2 quiet off-screen listener murmurs earn a place — sparse, woven through her speech (not parked only at the end), not after every clause, not a reverse-shot cutaway. Skip short callbacks and all sceptic-speaking beats.

Model prompt: never paste that judgment as a menu (uh-huh / oh? / mm). After you decide, write exact off-screen lines and where they sit relative to her words.

Example — agent chose two spots; prompt gets:

Copy code
[DIALOGUE]
Believer says exactly: "Those water-retention drops changed everything —
one salty meal prep day and I'd look puffy. Now I just don't."
Voice: <believer Voice lock verbatim>

[BACKCHANNEL]
Off-screen male partner only (not on camera). Same sceptic Voice lock, quieter than her.
After she says "changed everything" he says exactly: "Oh?"
After she says "I'd look puffy" he says exactly: "Really?"
No other reactions. Do not invent a second person in frame.

Step 2 — Two matching start frames

list_models({ type: "image" }) → generate_image for each speaker.

Still models (validated)

Do not default to Nano Banana / Krea / Ideogram / Recraft / MAI / Reve for podcast stills unless asked or the preferred trio fails.

These stills are shot / reverse-shot anchors: same room + camera grammar, opposed eyelines.

Hard rules

1. One person only. No foreground shoulder, OTS, interviewer blur, or second body.
2. Same camera grammar on both. Height, distance (chest-up), table edge, wall, lights — copy-paste. Only wardrobe + face + gesture + eyeline side change.
3. Same room word-for-word.
4. Flip the reverse shot. Believer frame-RIGHT ↔ sceptic frame-LEFT (or vice versa). Mirror mic/boom side with the subject.
5. Generate A, then B with referenceMediaRefs: [stillA] (use GPT Image 2 / Seedream if A was Grok): preserve room + camera height/distance — replace subject and flip eyeline.

Prompt pattern

Copy code
Podcast studio mid-shot, chest-up, eye-level camera, locked framing.
Single subject only — no other people, no foreground shoulder, no over-the-shoulder.
Wooden table edge low in frame, broadcast mic on boom in shot.
Background: <identical room description>.
Warm practicals, normal exposure, real room — not a film set.
Subject mid-gesture, mouth slightly open — not into lens.

Then only subject + eyeline differ:

Copy code
Believer: <wardrobe/face>. Eyeline: looking frame-RIGHT toward off-screen partner.
Sceptic: <wardrobe/face>. Eyeline: looking frame-LEFT (OPPOSITE of believer).
Preserve room and camera height/distance from reference — flip orientation only.

Negatives: no second person, no OTS, no interviewer, no looking into lens, no same-side eyeline as the other speaker.

Frame checklist

• One person only; no OTS / shoulder
• Matching height, distance, room
• Mic in front of speaker
• Off-axis eyeline; opposed across the pair
• Mid-gesture / mouth slightly open

folder: "PodcastAd/Stills"

Step 3 — Generate conversation beats

1. get_timeline — set delivery aspect (often 9:16) via set_project_settings before placing if needed.
2. list_models({ type: "video" }).
3. Prefer Flux 3 draft: true for dialogue naturalism (validated bakeoff). Also fine: seedance-2-mini, MiniMax H3 (hailuo-03). Lip-sync misses: Kling V3/O3 or Grok.
4. Final: Flux draft → generate_video({ enhanceDraftMediaRef }) only (FLUX Enhance; not upscale_media; needs canEnhanceDraft). Else re-gen on seedance-2 / H3.
5. Propose beats; wait for confirmation.
6. startFrameMediaRef = that speaker’s still. Seedance: framesAndReferencesExclusive — don’t combine first frame with reference image pools.

Voice lock (required)

Before any generate_video, write two fixed Voice lines with explicit accent + tone + pitch + pace. Paste verbatim every beat for that speaker. Name the accent the user wants (e.g. Australian, US California, UK).

Copy code
Believer Voice: <accent>, <pitch>, <tone>, <pace>, no dead air
Sceptic Voice: <accent>, <pitch>, <tone>, <pace>, no dead air

Example (only if the brief asks Aussie):

Copy code
Believer Voice: Australian woman, natural Aussie accent not exaggerated,
warm mid pitch, lightly upbeat certainty, brisk conversational pace, no dead air
Sceptic Voice: Australian man, natural Aussie accent not exaggerated,
medium-low pitch, dry sceptical tone, measured but snappy, no dead air

Never rewrite mid-ad. Optional: beat-1 audio as @Audio1 on later Seedance/H3 calls.

Prompt extras every beat

Copy code
[SCENE]
Animate from the start frame. Overheard — talks to off-screen partner, not the lens.
Locked camera; ambient room only.

[GEOGRAPHY]
Match the start frame: one subject, mic in shot. Hold the SAME off-axis eyeline
side as the still for the entire clip. Natural micro head turns only within that
side — never swing to camera, never flip to the opposite side.

[DIALOGUE]
<Speaker> says exactly: "<line>"
Voice: <paste that speaker's locked Voice line verbatim>

[BACKCHANNEL]   // believer long turns only, after agent decided exact lines
...exact off-screen lines + placements...

[ACTING]
Talk to the off-screen partner the whole time. Fill the duration with speech —
minimal trailing silence. No cheerleading. No hard sell to lens.

[CAMERA]
Hold the start-frame angle. No orbit, no new coverage.

[NEGATIVES]
No looking down the lens, no eye contact with camera, no turning to face camera,
no flipping eyeline, no second person, no OTS, no room/light change, no CTA
lower-third, no identity swap, no subtitles/watermark, no long pauses.

Regenerate that beat only on failure. folder: "PodcastAd/Beats"

Step 4 — Assemble + finish

Copy code
add_clips({
  entries: [
    { mediaRef: BEAT1, startFrame: 0 },
    { mediaRef: BEAT2, startFrame: END1 }
  ]
})

• remove_silence (tight: minimumPauseSeconds: 0.25, light speechPaddingSeconds) after assembly
• Product / before-after over mechanism beat: higher video track or apply_layout PIP
• Captions: add_captions — clean default unless asked
• Do not over-polish into “ad grade”

Turning one into thirty

Anti-patterns

• Writing believer first; sceptic has nothing to do
• Sceptic ends as brand ambassador
• Start frames or video beats staring at / turning to camera
• Same-side eyelines on both speakers
• Vague Voice lines (no accent + tone) — voices drift
• Vague backchannel menus in the model prompt (agent must pick exact lines)
• Reverse-shot cutaways for uh-huh / oh? (bake off-screen under her instead)
• OTS / foreground shoulder / second person in a “single” still
• Over-long durations leaving dead air
• Vague mechanism (“it just works”); hard CTA in dialogue
• One overloaded clip for the whole ad; skipping timeline assembly