Lock the voice first: a reference-role Seedance prompt that one-shots talking ads, and ElevenLabs' $100k fake-product ad contest (closes Nov 1)

@ZssBecker posted the technique on Oct 22 as a reply to another creator who was struggling with talking shots. It passed 760 likes, 2,000 bookmarks and 70k views within a day. His diagnosis: when the model makes up the speech itself, it adds or adjusts movement and words to fit the body motion. His fix is to make the audio first, attach it as @Audio1, quote the script word for word in the prompt, and state that the dialogue is reused verbatim (the same words, count, order and timing) and is the only speech in the clip.

The prompt itself is a long template split into labelled sections. Generation Goal is a one-paragraph shot description. No Words On Screen bans captions and UI. Reference Material Roles gives each input exactly one job: @Image1 is the person, @Image2 the location, @Image4 what appears on a laptop screen, @Video1 drives only facial expression and carries no audio, @Audio1 is the voice and full dialogue. The other sections are Dialogue, Subjects, Scale and Proportion, Camera, Picture Quality, Lighting, Performance, a timed Event Script ('around 8 seconds in…'), Audio and Maintain Consistency. A lot of it is negative direction: no wide eyes, no flicker, no stabilised glide, no cuts.

The reusable lessons are to make the audio the master track, say what each reference does not control, and script beats by timestamp. His example is a deliberately low-fi selfie ad. Use your own face or a consenting actor, don't put made-up results or dashboards on screen, and follow your ad platform's AI-disclosure rules.

Somewhere to try it: ElevenLabs' 'The Search' wants a new 16:9 ad, 15–45 s long, for a fictional product, with an original jingle made in ElevenCreative. Entries need the competition watermark and must be posted on two platforms (X, LinkedIn, Instagram, TikTok or YouTube). Prizes total $100k across 11 winners, with $50k for first. Submissions close Nov 1, 2026 at 11:59 PM PST, and winners are announced Nov 10. @maxxmalist posted a plan for using Claude Code to brainstorm and rank products and jingles before you build the entry in ElevenCreative. Not tested by us.