POV Hand Kiss · Open Skill 공개 스킬
POV Hand Kiss
POV Hand Kiss
Custom
1 0 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-10-13/v8exvqd7.jpg
# POV Hand Kiss
Turn **one uploaded character image** into a **5-second first-person POV video** of that same subject kissing the viewer's hand. There is no project state. Progress uses `write_todos` plus the canvas video.
This skill does **only** that shot: the **character in the photo** kisses the **POV viewer's hand**. Nothing else.
## When to use
- User uploads (or points to) a photo/illustration of a character and wants a POV hand-kiss video.
- User asks to regenerate, retake, or slightly restyle **the same** hand-kiss clip (slower, more tender, different hand, match a new photo).
- User says kiss / 亲吻 / POV / first-person **and** the target is the viewer's **hand**.
## When NOT to use
Stop and do **not** call `generate_video` if the user wants:
- Mouth-to-mouth, neck, cheek, or any non-hand kiss
- The POV viewer kissing the image subject (roles reversed)
- Hugs, dance, talking-head, multi-shot films, or any other video type
- A video with **no** uploaded/referenced character image
- NSFW / sexualized action (video generation does not support adult content)
- A subject who appears to be a minor
Politely say this skill only makes **POV hand-kiss** clips from an image, and ask for a suitable adult character photo if needed.
## Production stages (references/)
Do **not** read every file at once. Read only the file for the current stage with `read_skill_file`.
| Stage | File | Tools | When to read |
|---|---|---|---|
| 1. Analyze the image | `references/image_analysis.md` | `recognize_image` | User provided an image, or you must identify the subject before writing a prompt |
| 2. Write the kiss prompt + generate | `references/kiss_action_prompt.md` | `generate_video` | Subject is locked; time to assemble the English prompt and call `ref2v` |
## Collaboration
- If an image is already in the conversation, **do not wait**. Run Stage 1 then Stage 2 in the same flow.
- No image → ask the user to upload one character image. Do not invent a subject.
- Do **not** ask for confirmation between analysis and generation unless the user said to wait.
- On first run, `write_todos` with: `analyze` → `prompt` → `generate`. Keep only one item `in_progress`.
- After a successful video (or a hard stop), call `set_guidance` in the **same turn** as the final reply. Do not call other tools in that turn.
- Suggested next steps must stay inside this skill: regenerate, use another photo, make the same hand-kiss slower/more tender. Never suggest a lip kiss or a different genre.
- Reply to the user in **their language**. All `recognize_image` / `generate_video` prompts stay **English**.
- Final reply: short natural-language summary (who was detected, 5s POV hand-kiss). Do not paste long URLs or the full prompt unless the user asks.
## Hard rules
1. **Image is required.** `generate_video` must receive the user's character image in `input_urls`. Never text-to-video this shot.
2. **Identity lock.** The on-screen kisser is the uploaded subject. Do not swap face, hair, outfit, body type, or art style.
3. **Action lock.** The only action is a classic hand-kiss (back of the hand / knuckles). No licking, finger-sucking, or sexualized motion.
4. **POV lock.** Camera is the viewer's eyes looking at their own hand. Do not show the POV character's face.
5. **Duration lock.** Always `duration=5`. Do not generate other lengths.
6. **Channel lock.** Always `model='ref2v'`. Never pass a physical model name. Do not use `i2v` / `t2v` / `fl2v` for this skill.
7. Failures: report the tool error honestly. Never fabricate a video URL.
## Canvas
- `generate_video`: `name` like `{Subject} POV Hand Kiss` (or `POV Hand Kiss` if unnamed), `entity_type='final_videos'`.
- `recognize_image` does not go to the canvas.