---
name: design-delivery-reel
description: "Use when making a ~20-second 9:16 comic story reel with Seedance 2.5 starring the project person (e.g. designer Yujeong) and a real car model (e.g. Tesla Roadster), built from beats like change request → flying dispatch → print delivery; adding captions and exporting a Final; and handling feedback like garbled text by regenerating only the broken range into v2 or v3. Example triggers: make a Yujeong reel, design delivery, funny reel flying in a Tesla, 20-second reel, remake only the part where the text broke, regenerate just that part and splice it in, funny story reel with my character and a car, fix only the broken part."
---

# Design delivery story reel (generate → captions → partial regeneration)

When to use: only when making a comic short with several beats starring a person and a vehicle, or when applying feedback on a reel made that way. For work that follows a source video's motion, use the ramp-run-landing skill. Answer and report in the user's language. 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.

## 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. Who is the lead, and what are their details, like a job?
2. Which vehicle? Do you have a real reference photo? (default: if not, find and use a CC-licensed photo)
3. What story? A flow of 5–6 beats. With no idea yet, the agent suggests 2–3.
4. Any text visible on screen (phone, sign, document)? What exact wording? (default: screens with no text)
5. Captions in which language and tone? (default: one line per beat in the user's language)
6. Length and budget? (default: 20 s, show the estimate first)

## 0. Prep
- Check the person sheet with `includeImage=true`. Use character details like their job only if confirmed in the conversation or memory (e.g. Yujeong is a designer).
- **For a specific car model, always attach a real photo as a reference.** Text alone gives a different car (in v1 a Ferrari actually came out instead of the Tesla).
  - If the project has no photo, find a CC-licensed photo on Wikimedia Commons and bring it into the asset folder with `import_web_asset`.
  - Put the license and author in the file name. Example: `TeslaRoadster_reference_CC-BY_Jurvetson.jpg`

## 1. Storyboard (5–6 beats)
Example:
1. Driving home from work
2. Gets a "Just one tiny change~" request on the phone
3. The car flies off like a rocket
4. Races with the prints in hand
5. Delivers them through a high-rise window
6. Ends by leaving into the sky

**Decide in advance any shot where text is visible (phone, signs, documents) (project rule).**
- Write the screen content and the exact wording in the prompt. Example: phone screen shows only "Client" and "Just one tiny change~" with a blue "OK" button, nothing else. Use the user's own wording and language for these strings.
- Or make it a screen with no text (blank glowing screen, no readable text).
- The safest way is to make that shot's first-frame image separately first, check the text, then turn it into video (§4 steps 3–4).

## 2. Generate
- Model: Seedance 2.5
- References: [person sheet, car photo]
- configs: `1080p`, `durationSeconds 20`, `aspectRatio 9:16`, `generateAudio true`
- Prompt:
  - Write the beats in time order (first…, then…, finally…).
  - Point to the car through the reference photo, like "the exact car shown in the second reference photo (same body shape, color, headlights)".
  - Keep face and outfit, comic tone, and end with No logo.
- Cost: about 1,186 credits for 20 s. Get the estimate first and use that value as `maxCredits`.
- Check times: 0, 3, 7, 11, 14, 18, 20 s. Look at the car model, face, beat order and text.
- For a failed attempt, add the reason to the end of its name like `_try1_rejected(wrong-car)`, change its status to rejected and leave the reason in a comment.

## 3. Edit → Final
- Make a new video edit and put the clip on V1. Its sound links to A1 (`add_clip_to_timeline`).
- Put captions on V2 per beat with `add_text_clip`, leaving 0.1 s between captions. Style (`set_text_style`):
  - pretendard, weight 700, 68px, x 0.5, y 0.17, centered
  - outline #000000, width 0.12, text color white
  - only the last caption at 76px, y 0.2, color #FFE14D, two lines
- Export with `export_draft`. Name example: `Yujeong_design-delivery_reel_20s_Final.mp4`. Keep it in the same folder as the source and set its status to needs review.
- Leave a production note comment at 0 s. Include: model, references, beat structure, caption list, times checked, license source, credits.

## 4. Feedback: a problem in one range → regenerate only that range (the Final_v3 method)
Don't remake the whole thing. A 20 s regeneration is about 1,186 credits; a partial one is about 29 credits.
1. **Find the cut boundaries**: move through the editor with `seek_editor`, capture the canvas and find the problem cut's first and last frames. Example: frame 39 at 1.3 s – frame 75 at 2.5 s (30 fps).
2. **Save a clean source frame**
   - Hide the V2·V3 caption tracks and move to the middle of the cut (e.g. 1.9 s).
   - Save it with `Export → Save current frame → Add to project`.
   - Move the saved file to the asset folder. Name example: `phone-cut_source-frame_1.9s_reference.png`
   - Unhide the tracks again.
3. **Remake the first-frame image**
   - Model: GPT Image 2.5, with the one source frame as the reference (about 3 credits)
   - Prompt: "Use the attached reference frame as the base: keep exactly the same smartphone, hand, framing and lighting. Update ONLY the content of the phone screen: '9:41' / 'Client' / 'Just one tiny change~' / 'OK' button. Nothing else on the screen, no gibberish, text spelled exactly as given, no extra text. Save as PNG." Replace the quoted strings with the user's own wording and language.
   - Compare the result letter by letter. If it's wrong, write the wording more clearly and remake it only once.
4. **Turn it into a short video**
   - Model: Kling 3.0, first frame = the image above
   - configs: `1080p`, `3 s`, `9:16`, `generateAudio false` (about 26 credits)
   - Prompt: "Cinematic live-action footage, perfectly locked static camera, subtle natural motion only. The phone screen content stays perfectly still. No new text appears. One continuous shot with no cuts. First frame:"
5. **Replace**
   - Split the V1 source clip at the start and end of the problem range with `split_clip(includeLinked=false)`. Leave the A1 sound running as is.
   - Delete the middle piece.
   - Add the new clip, fit it to the range length (e.g. 1.2 s) with `trim_clip`, then place it with `move_clips` **in the same spot on V1**. On V3 it would cover the V2 captions.
6. **Check**
   - Use `get_timeline_layout` to confirm V1 is in the order [source | new cut | source].
   - Capture the new cut's start, middle and end, and the boundary frames before and after it.
7. **Wrap-up**
   - Export with a bumped version like `_Final_v3` and set its status to needs review.
   - Comment "replaced by v3" on the previous version and change it to closed.
   - Reply to the original feedback comment (range changed, method, cost, where to check). Look up the commentId first with `list_feedback`.
   - Leave a production note comment on the new version. Set the reference frame image to closed.
- Only when it must be hidden urgently, cover it with a white text card. If the user wants regeneration, follow 1–7 above.

## 5. Review tips and traps
- The project player can't jump to a chosen time. Capturing during playback only catches every 2–3 seconds, so a 1-second range gets missed. For short ranges, `seek_editor` in the editor and capture with the canvas ref. The capture can update one beat late, so retake it if in doubt.
- A newly uploaded file blocks capture for about 30 seconds until conversion finishes.
- Report any range of the exported file you didn't see directly as "needs a person to play and check".
- If you get the "The user operated the work screen" error, query again with `get_app_screen` and then continue.

## 6. Report
File name and folder / times checked / ranges not seen directly / actual credits charged (this task and cumulative) / rules to apply next time.