Halloween Smoke Magic Transformation · Open Skill 公開スキル

Halloween Smoke Magic Transformation

Halloween Smoke Magic Transformation 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。

@wwws

Custom

3 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/k5y8mqhb.jpg

# Halloween Smoke Magic Transformation

Converts one uploaded photo into an 8-second, one-take vertical short video: purple smoke rises from the floor and coils around the subject, a Halloween look (makeup + a miniature witch hat) appears piece by piece, the jack-o'-lanterns and candles behind them light up, and bats streak across the frame as the smoke clears.

The subject's identity, hairstyle, outfit shape and framing must survive the transformation — this is a costume/magic treatment on top of the user's photo, not a new character.

## When to use

- User uploads a photo and asks for a Halloween transformation / smoke magic / witch makeover video.
- User wants to preview the Halloween look as a still before animating it.
- User wants to adjust one beat of an already generated clip (smoke timing, hat, bats, lighting, pacing).

## Inputs and how they are used

| Input | Use in the prompts |
|---|---|
| Reference photo (required) | The identity and layout anchor. Analyse it first, then restate its concrete details (hair, face, outfit, framing, background) in every prompt — never replace them with a generic character. |
| Text (optional) | User's own wishes: mood, extra props, colour of the mist, "no hat", "keep my glasses", target duration inside the 8s envelope, etc. User text overrides the defaults below only where it is explicit. |

## Pipeline (references/)

Read only the file for the stage you are in — do not load them all at once.

| Stage | File | Tools | Read it when |
|---|---|---|---|
| Intake + identity lock | `references/workflow-stages.md` | `recognize_image` | Always, as step 1 |
| Look / style + prompt wording | `references/style-and-prompt.md` | `generate_image` | Before any image or video prompt is written |
| Optional Halloween keyframe | `references/workflow-stages.md` | `generate_image` (`i2i.hd.lite`) | User wants a still preview, or the photo needs the costume locked before animating |
| Detailed video script | `references/video-script.md` | — | Always, immediately before `generate_video` |
| Shot / motion / forbidden-content limits | `references/video-content-constraints.md` | `generate_video` | Always, together with the script step |

## Steps

1. If the skill was not read yet, `read_skill(name="halloween-smoke-magic-transformation")`. Then `write_todos` with the four stages (intake, look, script, video) and keep it updated with the full list each time.
2. `read_skill_file(name="halloween-smoke-magic-transformation", path="references/workflow-stages.md")` and follow its intake stage: run `recognize_image` on the uploaded photo, then write the explicit identity lock list (face, hair, skin, outfit silhouette, pose, framing, background) you will reuse in every prompt.
3. `read_skill_file(..., path="references/style-and-prompt.md")` and build this request's prompt from its Style / Subject / Layout / Must keep / Must avoid sections. Do not paste the reference verbatim; fill its templates with the user's photo details and the user's text.
4. Optional still preview — only when the user asks for a preview or the costume must be locked first: `generate_image(prompt=<keyframe prompt from style-and-prompt.md>, input_urls=[<user photo>], model='i2i.hd.lite', orientation='portrait', name="Halloween look preview", entity_type='characters')`. Show it, and if the user rejects it, fix the prompt before spending a video generation.
5. `read_skill_file(..., path="references/video-script.md")` and write a complete, request-specific English script for this upload: logline, 8-second beat map with per-beat shot size, camera move, subject action, VFX/look layer, lighting, and transition into the next beat. It must follow the user's text and the analysed photo — a template restated with no new detail is not acceptable.
6. `read_skill_file(..., path="references/video-content-constraints.md")` and fold its one-take rules, motion limits and forbidden-content list into the same prompt.
7. Compress the script into one continuous-sh