---
name: dance-character-swap
description: 'Use when making, from start to finish, a 9:16 short from one dance video (screen recordings from social apps included) and one photo of a person, replacing the person in the video with that person (or a fictional character) while keeping the dance moves and the original song timing. Covers the character sheet, looping and UI cleanup of the recording, the Seedance 2.5 swap, song sync and speed correction, export and review. Example triggers: swap the dancer with my character, replace the person in this dance video using a character sheet, keep the song timing, turn this dance video into Yujeong, make a challenge video with my character.'
---

# Dance video character swap (one character sheet + original song kept)

When to follow: only for requests to replace the person in a dance or performance video with the project's person (one photo) while keeping the original song. Ignore it for any other request. UI names below are given in English; if the app shows them in another language (for example Korean), find the item with the same meaning. Write replies, comments and reports in the user's language.

## Ask the user before starting
The point of this skill is not to remake the example but to make the user's own video. Before starting, ask the questions below together, in one message.
- Skip anything the request already answers.
- If the user says "you decide" or "use the defaults", go with the defaults in parentheses.
- If the user seems to have no idea yet, briefly suggest 2–3 options that differ from the example.
- Once answered, show a one-line summary of what was decided, then start.

1. Which dance video, and which part of it? (default: one full pass of the song)
2. Who should the dancer become? Point to a person photo or character.
3. If several people appear, which one changes? (default: the main dancer in the middle)
4. What outfit? Share any concept you have in mind. (default: the original dancer's outfit silhouette)
5. Zoom in so the face reads larger? (default: keep the original framing)

## 1. Rules
- Input: one dance video (Instagram or TikTok screen recordings work) + one photo of the person (a large, clear front-facing face; several angles in one image is even better). Do not ask anything beyond 'Ask the user before starting'; then proceed. Without an outfit request, follow the original dancer's outfit silhouette.
- The person must be the user, someone who consented, or a fictional character. If they look like a real celebrity, stop and say so. The original video may be someone else's post, so note in the report that rights must be checked before publishing outside.
- Output: 9:16, 1080×1920, 30fps, the original song as is. Generate video with sound off (`generateAudio` false).
- Resolution: make the final with Seedance 2.5 at **1080p**. A 720p generation is scaled up in the 1080 edit and looks soft (real feedback: "the quality is low").
- Choosing the video: if single-color stage light (red or blue) covers the face or the dancer is small on screen, the face stays hard to see after the swap. Where possible, pick a stretch or video where the face is bright and large.
- Credits (reference; the real figure comes from `estimate_ai_generation`): character sheet about 20–65 / Seedance 2.5 1080p 15 s about 1,000, 23 s about 1,600 / 720p 23 s about 650.

## 2. Steps
1. Register the steps with `update_task_scope`. Under the result folder, `create_folder`: 01_Input, 02_CharacterSheet, 03_CleanSource, 04_AISwap, 05_Final, 99_Previous_Rejected. Split file moves and renames into `apply_changes` calls of 10 or fewer (more gives a 429).
2. Analysis edit (1080×1920): put the video on V1 with `add_clips_to_timeline`. If no linked audio appears, the source has no sound (common with screen-recorded .mov files) → note it in the report. Open the editor through the sidebar `Edit` → `Video editor` (more reliable than navigating by address).
3. Finding the range: `inspect_app_media` in the project preview often cannot read the recording's codec. Look with `seek_editor` + `capture_app_screen` at 1-second steps, then 0.1-second steps. Range to use = from the post's first frame to the first point where it starts repeating, or until a menu or swipe covers it (e.g. 0.0–23.0 s). This range is one full pass of the song.
4. Cleanup: cut with `trim_clip` (includeLinked) and zoom in with `set_layer_transform` to push the UI off screen (Instagram Reels scale 1.25 · y 0.525, recordings with margins scale 1.3 · y 0.526). Check start, middle and end captures for leftover buttons, account names or captions. `export_draft` (startSeconds · endSeconds) → 03_CleanSource, name `<prefix>_<video>_clean_<length>s_<song|silent>`.
5. Character sheet (Nano Banana Pro `gemini-3-pro-image`, reference [person photo]): the §6-A prompt. Put the result next to the photo and compare face shape, eyes, nose, mouth and hair. If the face turns round or face gems, earrings or glasses appear, it failed → remake with references [person photo, previous sheet] adding "face exactly like the photo, no decorations". Only a passing sheet goes into 02_CharacterSheet.
6. Swap generation (Seedance 2.5): attachmentIds [clean source, character sheet, person photo] — **always include the original photo too** (with only the sheet the face goes soft). configs: resolution 1080p, aspectRatio 9:16, durationSeconds = clean source length rounded up, generateAudio false. Prompt §6-B. Estimate with the same attachments and settings → `start_ai_generation`. 23 s at 1080p takes about 13 minutes. When done, `rename_file` `<prefix>_<video>_swap_seedance25_1080p_vN` → 04_AISwap.
7. Assembly: `add_clip_to_timeline` the generated clip on the track above the original (V2/V3) at the same start time. 1080p is scale 1.0 (720p would be 1.5). Keep the song as the original clip's audio (A1).
8. Speed correction (required): put the generated clip's `sourceDuration` from `get_editor_context` ÷ the clean source length into `set_clip_speed` (e.g. 22.708 ÷ 23.0 = 0.9873). After correcting, compare the pose at the end point with the original.
9. Bigger face (optional): if the dancer is small, `set_layer_transform` scale 1.1–1.15, y 0.52–0.53. The toes get cut a little and the face gets larger.
10. Repeated start check: if the same move appears twice in the first 1–3 seconds of the generated clip or the loop overlaps, cut that much off the front (real case: about 2.2 s cut from the front of 23.0 s for a 20.8 s final). Cut video and song together.
11. `export_draft` → 05_Final, name `<prefix>_<video>_<person>swap_Final_1080p` → `set_file_status` review, and `create_comment` a task comment at 0 s (range, attachments, speed, scale).

## 3. Checks
- The first-frame pose matches the original's first frame.
- Middle and end: toggle the generated layer with `set_layer_opacity` 0 ↔ 1 and compare with the original at the same moment (arm raises, crossings, pause beats).
- Face: reads as the same person as the photo; leftover original props (sunglasses etc.), distorted hands, outfit changes, leftover UI or watermarks.
- If the result file's codec cannot be read, use editor captures instead and report that song sync needs a human listen.

## 4. When it fails
- Low quality: if it was 720p, redo at 1080p. Zooming in the edit does not fix it.
- Doesn't look like the person: ① remake the sheet comparing its face with the photo ② add the original photo to the attachments ③ add a face description + "face well exposed and in sharp focus" to the prompt ④ if lighting still covers the face, remake with a sentence changing the light to neutral, or change the video.
- Moves are off: first do the speed correction in step 8. If still off, generate in pieces of about 10 seconds and join them.
- If the same defect appears twice, stop editing the prompt and change the video, range or model.

## 5. Report format
Result file (name, folder, length) → range used (seconds in the original) → attachments and settings → speed and scale → moments checked → remaining defects (for a person to look at) → credits (estimate and actual charge kept separate).

## 6. Prompts
A. Character sheet: "Character reference sheet of ONE person. Face and hair must exactly match the person in reference 1: same face shape, eyes, nose, lips, skin and hairstyle. Natural light makeup only; no face gems, no glitter, no earrings, no glasses unless the photo has them. Outfit: <outfit>. 16:9, plain light grey studio background, photorealistic, soft even light. Top row: three large head-and-shoulders close-ups (front neutral, three-quarter smiling, profile) where the face fills each panel. Bottom row: full body front, three-quarter, back, and one dance pose. Same person and outfit in every panel. No text, no labels, no watermark."
B. Swap: "Edit the attached video (first reference). Replace only the dancer with the person shown in the character sheet (second reference) and the identity photo (third reference). Her face must be clearly recognizable as the person in the third reference in every frame: <face and hair description>. No glasses, no sunglasses, nothing covering the eyes. Outfit from the character sheet: <outfit>. Keep everything else exactly as in the original video, frame by frame: the same choreography, every movement with the same timing and rhythm, the same camera movement and framing, the same set, lighting and crowd. Let the face stay well exposed and in sharp focus. Same duration, starts on the same first frame, one continuous shot with no cuts. Sharp, high-detail cinematic live-action footage. No text, no captions, no app interface, no watermark."

Change log: v1.0 2026-10-11 first version / v1.1 generated clip scale, speed correction, opacity comparison / v2.0 2026-10-11 generalized to one-photo input, 1080p default, original photo added to attachments, sheet face comparison and remake, folder structure, bigger face, cutting a repeated start, lighting and size caution, how to open the editor, apply_changes in batches of 10 / 2026-10-19 added questions before starting