ugc-multicut-script

Writes the directorial prompt that drives the single Seedance generation.

Inputs

• PROFILE (product-profile.md), BRAND_FACTS (voice/positioning), DEST, optional DURATION (default 15, max 15).

Steps

1. Map three cuts onto the timeline (e.g. 0–4s Hook, 4–10s Setup/Action, 10–15s Recommendation), each matching its storyboard panel.
2. For each cut, specify:
   • camera behavior (handheld micro-shake for selfie hook vs locked-off for macro),
   • the physical interaction (the exact usage step from PROFILE — sanitizing, pouring, applying, gesturing),
   • the spoken monologue line (natural creator voice, brand facts woven in, NO em-dashes; write brand names / unusual words PHONETICALLY — see the VO Pronunciation Playbook below),
   • sound cues (ambient room tone, product sfx).

3. Write DEST/script.md:

   # UGC Script — <slug> (<DURATION>s, 3 cuts)
   
   ## Cut 1 — Hook (0–Xs, Tight, handheld micro-shake)
   - Action: ...
   - VO: "..."
   - SFX: ...
   
   ## Cut 2 — Setup/Action (X–Ys, Macro, locked-off)
   - Action: <exact usage step>
   - VO: "..."
   - SFX: ...
   
   ## Cut 3 — Recommendation (Y–<DURATION>s, Wide, handheld)
   - Action: ...
   - VO: "..."
   - SFX: ...
   
   ## SEEDANCE_PROMPT
   <one dense paragraph concatenating the three cuts as a continuous motion+dialogue description — this exact block is what the video step passes to seedance_2_0. OPEN with a one-line voice spec (e.g. "The voiceover is a calm, credible late-20s female narrator speaking clear American English. No subtitles, no on-screen captions."). Introduce each spoken line with `she says:` / `he says:` and put the line in quotes. End the block with "No subtitles.">

Output

DEST/script.md containing a SEEDANCE_PROMPT block.

Constraints

• Total runtime ≤ 15s (Seedance hard cap).
• Monologue must be speakable in the allotted seconds (~2.5 words/sec). (~32 words across 15s validated comfortable.)
• No em-dashes / en-dashes anywhere in the VO lines.

VO Pronunciation Playbook (validated 2026-06-22, Seedance 2.0 native audio)

Seedance synthesizes the VO from the literal text in the prompt, so spelling drives pronunciation. Two failure modes seen live:

• Letter-by-letter spell-out: a non-word brand like nustandardlabs is read "N-U-standard-labs".
• Mangling: the naive fix New Standard Labs came out "newstandard labbers" (words fused + a schwa turned "Labs" into "labbers").

Rules that fixed it:

1. Respell phonetically with clear word boundaries. nustandardlabs -> Noo Standard Labz. Use "z" for voiced plural endings ("Labz" not "Labs") to stop schwa insertion. Test it by reading your spelling aloud as a stranger would.
2. Explicit dialogue form: she says: "..." (not saying, "...") — uses your exact words.
3. Append No subtitles. or the model burns auto-captions into the video.
4. Specify the voice once at the top of SEEDANCE_PROMPT: tone + age + gender + language.
5. Only a human can verify the audio. Render 720p first (cheap; audio is resolution-independent), confirm pronunciation, keep that take. Never pay for 1080p just to check audio.

Keep the WRITTEN brand spelling correct everywhere else (profile, caption, on-screen label) — ONLY the spoken VO quote gets the phonetic respelling.

Known-fix lexicon (drop these spellings straight into the spoken VO quote)

When a word with an aɪ.əl glide (vial / trial / dial / phial) misfires, anchor it to a real one-syllable word the TTS already knows rather than hyphenating syllables.