Fingertip Touch · Open Skill 公開スキル

Fingertip Touch

Fingertip Touch は Digen 上の公開 Open Skill です。開くと、このワークフローが何を作るか分かります。

@JackRhino

Roleplay

0 0 1

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-21/ssczaef5.jpg

# Fingertip Touch

Turn one uploaded reference photo into the signature "fingertip touch" frame — the subject leans in, one index finger reaches out and almost presses the lens, the fingertip blurs in the very front while the eyes stay sharp — and optionally animate that frame into a short, single-take clip. There is no structured project state; the reference photo, the generated still and the generated clip are the only carriers of continuity.

## When to use

- The user uploads a photo and asks for a fingertip-touch image, a "point at the camera" shot, or a hand-reaching-to-lens portrait.
- The user uploads a photo and asks for a short video / clip / animation of that subject reaching toward the camera.
- The user already has a fingertip-touch still or clip in this conversation and wants it re-rendered, re-animated, or tweaked (pose, framing, light, motion, duration).
- Do not use for unrelated image or video requests that have no "reaching fingertip toward the lens" idea.

## Inputs and outputs

| Input | Output |
|---|---|
| One reference photo of a subject (required) | One still image in the fingertip-touch look (`generate_image`) |
| Optional: user's own wording about mood, wardrobe, setting, duration | Optional: one short clip animating that still (`generate_video`) |

How to use the reference: it is the **identity anchor**, not a style suggestion. The face, hair, body proportions and any distinctive feature must survive the render. Never describe the person from scratch in the prompt as if the photo did not exist; call `generate_image` with the photo in `input_urls` so the model edits it instead of inventing a new person.

## Steps

1. **Read the request and set the scope.** Decide whether the user wants the still only, the still plus the clip, or a revision of something already generated. If more than one output is expected, call `write_todos` with the steps (analyse → still → clip) and keep exactly one item `in_progress`.

2. **Analyse the reference.** If the photo is ambiguous (multiple people, unclear framing, unusual lighting, dominant background props), call `recognize_image` with a prompt asking for subject identity traits, hairstyle, wardrobe, framing, camera angle and light direction. Note, in one short paragraph, what must be preserved.

3. **Read the visual rules.** Call `read_skill_file` with `path="references/style-and-prompt.md"`. Follow its style, subject, layout, must-keep and must-avoid sections, and build the still prompt from its template and fragment library. Do not write the prompt from memory.

4. **Generate the still.** Call `generate_image`:
   - `prompt`: English, assembled per `references/style-and-prompt.md` (style → identity anchors → fingertip gesture → layout/depth layering → light → quality).
   - `model='i2i.hd.lite'` (reference-image editing; the uploaded photo goes in `input_urls`).
   - `input_urls=[reference photo asset id or uri]`.
   - `orientation='portrait'` by default; only switch to `landscape` or `square` if the user explicitly asks for another ratio or the reference is unusable in 9:16.
   - `name`: short canvas title such as `Fingertip Touch Still`.
   - `entity_type='characters'`.
   - Leave `resolution` unset unless the user asks for a specific quality.

5. **Gate the still before animating.** Call `read_skill_file` with `path="references/step-details.md"` and run its still-review checklist: identity preserved, single hand with a correct finger count, fingertip clearly closest to the lens, eyes sharp, background clean. If it fails, fix the prompt per the retry rules there and regenerate once or twice; never animate a broken still.

6. **Confirm the video intent.** If the user only wanted the image, stop here and reply. If the user wants a clip (or already asked for one up front), continue.

7. **Write this request's script.** Call `read_skill_file` with `path="references/video-script.md"`, then write a complete, request-specific English script for this shot: logline, durat