---
name: drama-character-swap-shorts
description: "Procedure that takes only one live-action source video (drama, variety show, interview etc.) and one character photo (or character sheet), replaces one chosen person in the video with that character, and keeps the original lines, lip sync, other people and cut structure, from start to finish through a 16:9 final and a 9:16 short. Includes applying feedback. Example triggers: replace the woman in this video with Yujeong, swap the person keeping the lip sync, 100% sync and change only the face, the lip sync is off so fix it from the feedback, make it a short, here are the original and a photo so make it, replace one person in this clip with my character keeping lip sync, then make a 9:16 short."
---

# Live-action character swap → feedback → short (from the original + a character only)

When to follow: requests to replace **only one chosen person** in a live-action source video with a character while keeping the original sound, other people and cut structure. For dance videos use dance-character-swap; for vehicle videos use vehicle-character-swap. 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 source video, and who to replace? Distinguish them by clothes, hair or position.
2. Which character should they become? Point to a sheet or photo.
3. Only the 16:9 final, or a 9:16 short too? (default: both)
4. Add a title or captions in the short's black bands? (default: none)
5. Budget? (default: show the total estimate first, then proceed)

## 0. Inputs and defaults
- Required inputs: ① the source video file ② the character image (a sheet including a front-facing face is recommended) ③ who to replace (distinguished by clothes, hair, position).
- Anything not answered in 'Ask the user before starting' uses the defaults below. If the person to replace could be read as more than one person on screen, confirm again.
  - Budget: if the user hasn't set one, show the total estimate first and proceed (in auto-approve mode, proceed at 1,000 or less).
  - Video model: Kling 3.0 Omni, 720p, 4 s (5–8 s for long cuts), 16:9, generateAudio false.
  - Image model: Nano Banana Pro.
  - Short: 1080×1920, centered, black above and below.
- Never modify the original edit or the original file. Leave results only as new files (_v1, _v2 …).
- If the source carries rights marks such as a broadcaster logo, note at the end of the report that rights need to be checked.
- Every generation: estimate → generate → record the actual charge. Keep the running total in the project memory.

## 1. Preparation (0 credits)
1. Make two folders: `{job}_AIGenerated` (references and generations), `{job}_Final` (exported results).
2. Make a **new 16:9 edit** (1920×1080, 30fps) and put the source on V1 (video) and A1 (sound) from 0 s.
3. Check the source's frame shape with a capture and record it.
   - Is the whole frame picture, or a 4:3 picture + black side bars (used for the short's scale)?
   - Also record the source's scale in the editor (e.g. 1.5).
4. Make a **cut table**. Scan the source with captures every 0.5 s and list only the cuts where the person to replace appears.

   | Cut | Start–end (s, 30fps frame boundaries) | Person size | Line | Key moments (mouth shapes, hands, props) |
   |---|---|---|---|---|

   - Confirm cut boundaries by capturing one frame (0.0333 s) before and after.
   - Pick three or more key moments per cut: laugh, round "oh", wide-open mouth, wide eyes, hand or chopstick position.

## 2. Per-cut replacement — proven method (lip sync first)
Take one cut all the way through before moving to the next.

### 2-1. Make a real-time reference clip (0 credits)
1. Put the source in again in an empty time range of the same 16:9 edit (e.g. after the source's end plus margin, after 40 s).
2. Trim that clip to leave only the cut's range. **Keep the speed at 1.0** (a reference stretched slower throws off mouth timing).
3. If the cut is shorter than the generation length (4 s), save the cut's last frame as a PNG with `Save current frame` and place it right after to fill the remaining time.
4. Export only that range with `export_draft` (start and end set) into the `_AIGenerated` folder.
   - Name example: `{job}_refcutNN_realtime_{start}-{end}s+hold{x}s.mp4`
   - Record the reference length R (e.g. 4.0 s).
5. After exporting, delete the working clip and PNG from the edit.

### 2-2. Save the first frame (0 credits)
- Move to the cut's start → Export → `Save current frame` → the `_AIGenerated` folder.
- Hide other clips on upper tracks with opacity 0 so you save **only the original picture**.

### 2-3. Face composite on the first frame (Nano Banana Pro, attachments: source frame + character)
Prompt skeleton:
"Edit image 1. Replace ONLY the [person description] woman's identity with the person in image 2 ([character traits: age, hair, face shape, skin, eyes, lips]). Her face must clearly be image 2's person. Keep EVERYTHING else identical to image 1: same 16:9 framing and black side bars, same crop, same head tilt, same expression and eye direction, same hands and props at the same place, same clothes, same hair length and silhouette, same background, same lighting and broadcast image quality, same logo unchanged. Do not add or remove anything."
- Open the result and check: is the face the character, and are pose, hands, props and logo the same as the original? If not, fix the prompt and remake only once.

### 2-4. Video edit generation (Kling 3.0 Omni)
- Attachments: [real-time reference clip, composited first frame, character image]
- Settings: resolution 720p, durationSeconds = reference length, generateAudio false, aspectRatio 16:9
- Prompt skeleton:
"Edit the input video so the [person description] becomes the person shown in the first image (edited first frame) and the character sheet. Use that face for EVERY frame; the original actor's face must never appear. Keep everything else from the input video unchanged and frame-accurate: identical timing, head movement, blinks, hand and prop motion, and identical mouth shapes at identical moments so the lips stay in sync with the original dialogue. Same framing, black bars, background, lighting, clothes and logo. One continuous shot, no cuts, no added motion."
- Estimate → generate → poll the status until done (never repeat the same request).

### When it fails (if the same method fails twice, change the method)
| Symptom | Cause | Fix |
|---|---|---|
| Mouth shapes ahead of or behind the line | Slowed-down reference, Seedance keyframe generation | Real-time reference at 1.0x + Kling Omni edit |
| Lips and motion match but the original actor's face remains | Only the character image was given to the Kling edit | Add the 2-3 composited first frame too |
| The original face comes back midway | The reference face leaks in | First-frame composite + stress "never the original actor" in the prompt |
| Other people or fake cuts slip in | The reference includes a cut boundary | Shrink the reference range inside the cut and fill with the last frame |

## 3. Placement (theoretically exact)
1. Put the generated result on **V3** at the cut's start time.
2. Read the actual length D from `get_editor_context` sourceDuration (it differs every time, like 4.033 or 4.042 s).
3. Speed = D ÷ R (e.g. 4.0417 ÷ 4.0 = 1.0104). Enter it to 6 decimal places.
4. Trim the end to the cut's end time.
5. Scale: 720p result → 1.5 in a 1920×1080 edit. Use the same scale and position as the source clip.

## 4. Cut review (required for every cut)
1. Move to each key moment → capture (swapped).
2. Swapped clip opacity 0 → capture at the same moment (original) → back to 1.
3. Pass criteria: the face is the character, and mouth shape, head position, hands and prop angles match the original. No jump one frame before or after the cut.
4. When it passes, move it to the same range on V2. If there is a previous clip, delete it and leave V3 empty. Add the outcome to the generated file's name (`_adopted`, `_faceNotReplaced` etc.).

## 5. 16:9 final export
1. When every cut passes, export `{job}_final_v1.mp4` into the `_Final` folder (range 0 – source length).
2. Put the result file on an empty V track for a moment (lower tracks at opacity 0) and capture one key moment per cut plus one unreplaced cut. After checking, delete it and restore.
3. Leave the status at "Needs review". Only a person decides "Approved".

## 6. Applying feedback
1. Read the comment's time range, content and drawing.
2. Remake only that cut with steps 2–4. Do not touch the other cuts.
3. Export as `_v{n+1}` and review with step 5.
4. Reply to the original comment: "[Agent] Applied — what was changed and how / moments compared with the original / where to check: file name mm:ss–mm:ss"
5. The new version's status is "Needs review".

## 7. Short (9:16, black above and below, centered) — 0 credits
1. Make a **new edit** at 9:16 (1080×1920). Leave the 16:9 edit as is.
2. Put the latest 16:9 final file on V1 from 0 s (sound on A1).
3. Scale (in this editor, scale 1 = the source's pixel size as is):
   - Whole frame is a 16:9 picture: 1080 ÷ 1920 = **0.5625**
   - 4:3 picture + black side bars inside 16:9: 1080 ÷ 1440 = **0.75** (the bars fall outside the frame)
   - x 0.5, y 0.5
4. If the exported 16:9 file is a few frames longer than the original length, trim the end to the original length (video and sound together).
5. Capture a wide cut, a swapped cut and a cut showing the logo to check nothing is cut off.
6. Export `{job}_short_9x16_v1.mp4` into the `_Final` folder. Put the result file on V2 for a moment with V1 hidden at 0 to check, then delete it and restore V1. Status "Needs review".
7. Ask the user whether to add a title or captions in the black bands above and below; if so, add them as editor text (never draw lettering with AI).

## 8. Report format
- Results: 16:9 final file, short file name, folder and status
- Checks: per-cut moments compared with the original and the outcome (table)
- Cost: model, outcome and actual charged credits per attempt, running total and budget
- What a person should check: play at real speed with sound to check lip sync, cut start frames, rights to the source
- Project memory: folder IDs, edit IDs (16:9 and 9:16), adopted file IDs, per-cut speed and scale, running credit total