What the fixed skills did, as tools you call.

Script in quotes in the prompt; seconds from the word count (about 2.3 per second); refs = a portrait if the face must persist, addressed as @image1.

Poll with wait:true until completed; hand over the URL.

The product photo becomes an https URL.

"... holding THIS product up to the lens, label facing camera ..." with the product URL (and a portrait, if any) in refs.

The frame URL in refs as @image1, the pitch in quotes. The product stays the product.

The still: a portrait, a product shot, a scene. Or upload_image a photo the user has.

That URL as first_frame; the prompt says what moves and what is said. Leave aspect out (adaptive follows the frame). No refs in this call: frames and refs cannot be mixed on Seedance.

Poll until completed. The clip opens on the exact still.

Twice: the opening still and the closing still, same person and setting wording, a different pose or state.

first_frame = the opening still, last_frame = the closing still; the prompt describes the move between them. The model animates from one to the other.

With video_refs the quote adds the reference clip seconds at the same rate (measured at submit); price it first.

The clip in video_refs (https mp4/mov), a portrait in refs, and a prompt like "@image1 performs the same move as @video1, in a bright kitchen ...". Describe the NEW clip; never write edit or extend wording.

Poll until completed.

"Extreme close-up, face fills the frame, one exaggerated bug-eyed shock held straight into the lens, slow lean-in, no speech"; audio: false; seconds: 5; a portrait in refs (@image1) so it is the same face every time.

Burn the caption in the editor, or ask for a static caption in the prompt. Volume: same prompt + same refs, N calls, N performances.

The narration, a named voice, a tone.

Lay it over the footage in the editor. Muxing external video is not on this surface; the fixed make_subtitles and make_ugc REST routes still exist for that. The mp3 URL also works as an audio_refs entry for generate_video.

Once: a clean head-and-shoulders portrait.

Every clip: the same portrait URL in refs as @image1, the same person/setting wording, a different script. The reference is what keeps the face.

Saved characters from the dashboard appear here; their character_sheet_url works the same way in refs.

Both portraits in refs and a prompt that names who says what: "@image1 and @image2 on a couch. @image1 says: ... @image2 laughs and says: ...". Keep it to 10 to 15 s per exchange.

Price it first: model "seedance-2.5", seconds up to 10, the quality the user asked for.

One take, a strong reference. In image mode leave aspect out (adaptive only). Expect a much longer render than seedance-2.0 (12 to 25 minutes per clip; plan the wait).