---
name: youvico-underthefan-shorts
description: 'Use when making, from start to finish, a 9:16 fashion short (#underthefan) where the person from a single photo lies on a carpet directly under a ceiling fan (top-down view) and the outfit (look) changes along the blade each time a steadily turning fan blade passes over the body. All it needs is one photo of the person and a list of looks. Example triggers: under the fan reel, ceiling fan outfit change short, make an under-the-fan video from my photo, outfit transition reel, a short where my outfit changes as the fan passes, a video where the fan blade changes the clothes.'
---

# Under the fan outfit-change short (from one photo)

When to follow: only when making a vertical short where a person lies still in a fixed top-down view and the outfit changes as a fan blade passes over the body. 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. Who is it for? Point to the person photo in the project.
2. How many looks, and which outfits? A season or concept is enough. (default: the default 4 looks)
3. What background? Carpet and sofa colors, room mood, time of day (default: beige carpet, cream sofa, daylight)
4. What feel for poses and expressions? (default: still or cute poses)
5. Include the intro where they roll off the sofa? (default: yes)
6. What kind of music? (default: upbeat indie pop, no vocals)

## 1. Rules
- Input: one photo of the person (a clear front-facing, full-body photo works best). Without a look list, use the default 4 looks: ① shirt jacket + jeans ② field jacket + beanie + wide denim ③ shearling jacket + black turtleneck and slacks ④ pleated mini skirt + printed crop tee. Do not ask anything beyond 'Ask the user before starting'; then proceed.
- Use only photos of the user or of people who consented. If the person looks like a celebrity or public figure, do not proceed and say so.
- Output: 9:16, 1080×1920, 30fps. 4 looks + intro = about 16 seconds.
- Never make the fan with AI video. Fan = transparent PNG + rotation keyframes; person = Kling clips without a fan.
- The outfit change is a wipe that follows the blade. If the whole person changes in one frame, it failed. Do not use plain cuts or dip-to-black.
- The person stays still or does only cute poses (V sign, cheek heart, head tilt). No touching props.
- The fan stays visible until the very last frame (no opacity fade).
- Credits (reference; the real figure comes from `estimate_ai_generation`): images (number of looks + 2) × about 40, Kling (number of looks + 1) × about 65, one BGM. 4 looks ≈ 240 + 325 + BGM.

## 2. Timing plan (decide before generating)
- Rotation: 0–T0 s still at 30° (intro) → T0–T0+0.5 s ease-in to 20° → after that decrease at 40.1° per second (negative direction; blades pass from right to left).
- Time when blade i (relative angle aᵢ, reference blade 0°) covers the body: tᵢ = (T0 + 0.5) + (20 + aᵢ) / 40.1.
- Reference values (T0 = 6.4, blades at 0/128.4/231.6/360°): 7.40 / 10.60 / 13.17 / 16.38 s. With a new fan PNG, measure aᵢ again from captures (§3-6).
- Segments: intro and look 1 0–t₁, look 2 t₁–t₂, look 3 t₂–t₃, look 4 t₃–end (end = t₄ − 0.5, e.g. 15.9). Without the intro, T0 = 1.0.
- Source length needed per clip = segment length + 0.6 s before and after (for the 1.2 s wipe). A 5-second Kling clip is enough.

## 3. Steps (1–5: assets)
1. Register the steps with `update_task_scope`. Find the photo with `list_project_files`; if the user gives a local path, upload it with `upload_local_file`. Folders: Keyframes / Video / Final output / Rejected (`create_folder`).
2. K1 (GPT Image 2.5, reference [photo]): "Vertical 9:16 realistic photo shot from directly above (camera on the ceiling, perfectly top-down, no perspective tilt). The person from image 1 lies on their back on a plain beige carpet, full body centered, head at about 15% from the top, feet at about 85%, arms relaxed at the sides. A cream curved modular sofa runs along the left edge. Soft even daylight, no harsh shadows. Wearing <look 1>. Exactly the same face and identity as image 1. No ceiling fan, nothing above the person, no text." Check face, full body and head/feet positions with a capture. This K1 is the composition reference for everything after it.
3. K2…Kn (references [K1, photo]): "Keep exactly the same camera, carpet, sofa, lighting, body position, body size and framing as image 1. Change only the outfit to <look N> and the pose to a cute pose (V-sign near the cheek, smiling). Same face as image 2. No ceiling fan, no text." Capture each next to K1 and check that head and feet positions match within 3% of the canvas height. If one is off, remake only that image.
4. (Intro, optional) S0 (references [K1, photo]): "Keep exactly the same camera, carpet, sofa and framing as image 1. Same person and outfit, now lying on the cream sofa on the left, about to roll off onto the carpet. No ceiling fan." If the composition differs, remake once and move the failed file to the Rejected folder.
5. Fan PNG (GPT Image 2.5, no reference): "Square 1:1, fully transparent background, a ceiling fan seen from directly above, hub exactly centered, three identical wide taupe blades, one pointing straight up, brushed metal hub, no shadow, no ceiling." If one already exists in the project, reuse it.

## 3. Steps (6–13: editing)
6. Edit (new video edit, 1080×1920): put the fan PNG on V3 with `add_clip_to_timeline`, `set_layer_transform` scale 1.8, x 0.5, y 1.0. Capture to check that at 0° rotation a blade covers head to feet. Change the rotation to find the angle aᵢ at which each other blade is vertical, and measure the hub position at 0°/120°/240° to get the pivot distance r (px) (reference example: r≈203).
7. Keyframes: with `set_keyframes`, set rotation (the 3 segments of §2) and x·y hub correction (hold at the 30° position during the still segment; during rotation every 15°: x = 0.5 + (r/1080)·sinθ, y = y0 − (r/1920)·cosθ; reference example: 0.5+0.188sinθ, 1.026−0.1057cosθ). Do not add opacity keyframes. Fan clip length = to the end of the video.
8. Person clips (Kling 3.0, 1080p, 5 s, 9:16, `generateAudio` false, first frame [Kn]): "Use the attached image as the exact first frame. Locked top-down camera, no camera movement, no zoom. The person stays mostly still and does cute poses: smiles at the camera, keeps the V-sign near the cheek, gentle head tilt, playful blink. Hair moves slightly as if a ceiling fan breeze blows from above. Hands never touch any object. Same outfit, framing and lighting throughout. No ceiling fan visible, nothing enters the frame, no text." For the intro clip use first frame [S0] with "rolls off the sofa and lies on the carpet in the same position as the person in image 2" and references [S0, K1]. Start all clips at the same time and wait (`get_ai_generation_status`).
9. BGM (Seedance 2.5 or 2.0 Mini, 480p, length = video end + at least 1 s, sound on): music-led, "upbeat indie pop, no vocals". Do not stretch it with speed changes.
10. Placement: put each clip in an empty spot (after 80 s) and pick a good stretch with `seek_editor` + canvas capture. `trim_clip` so that 0.6 s before and after the stretch remain inside the source → `move_clips` into the §2 segments.
11. Transitions: from look 2 on, `add_transition` on each clip (edge start, kind wipe, direction left, 1.2 s, alignment center). If rotation runs in the positive direction, use direction right.
12. Sound: BGM on A2 with `add_clip_to_timeline` audioOnly, `trim_clip` to the end. Split the last 1.5 s into 0.5 s pieces with `split_clip`, `set_clip_volume` 0.6/0.32/0.1.
13. `export_draft` (0–end, Final output folder) → `set_file_status` review. File name example: `<name>_underthefan_v1.mp4`. Give generated files names that describe their content with `rename_file`.

## 4. Checks
- Capture each tᵢ at −0.15 s and +0.3 s: left of the blade is the previous look, right of it the new look, and the boundary never shows outside the blade. That passes.
- Middle of each look segment: same face as the photo, requested outfit, no prop touching, no fan.
- The fan is still visible in the last 0.1 s. Hub position difference between 0°/120°/240° is within 3% of the canvas width.
- If measuring in the result file viewer is blocked by "could not decode the codec", use editor `seek_editor` + canvas capture instead, and report that the sound needs a human listen.

## 5. When it fails
- Black screen at a transition: a dip-to-black left at the clip end (it comes along when trimming) → delete the transition in the transitions tab (needs app confirmation).
- Boundary shows outside the blade: change the wipe length to 1.0–1.6 s or move the boundary time by ±2–4 frames. If body size differs between looks, remake from that Kn.
- Face differs from the photo: remake only K1 (reference [photo] first, "exactly the same face"). After two failures, ask for a sharper photo.
- Kling shows a fan, objects or prop touching: strengthen the forbidding sentence and remake once only; for the rest, use only the good stretches.
- Replying to feedback: before `reply_comment`, first query that file's `list_feedback`.

## 6. Report format
Result (file name, folder, length), segment table (time, look, transition), moments checked, what a person should check (sound, boundaries), total credit estimate (reserved; the actual charge is the same or lower).

Change log: v1.0 2026-10-05 first/last-frame wipe / v1.1 fan PNG + rotation split / v2.0 one-photo input, wipe follows the blade, no fan fade or dip-to-black, still and cute poses / v2.1 2026-10-05 timing plan moved before generation (cover-time formula), still → ease-in → steady rotation in 3 segments, K1 composition reference check, optional sofa intro step, 0.6 s wipe margin, BGM generated to length, default 4 looks, credit formula / 2026-10-19 added questions before starting
