Floating Candle · مهارة Open Skill عامة
Floating Candle
Floating Candle هي Open Skill عامة على Digen. افتحها لتعرف ماذا ينشئ هذا التدفق.
Custom
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/67mcm23t.jpg
# Floating Candle
Turn a single reference image the user uploads into a Halloween "floating candle / jack-o'-lantern" short video: reference-to-video, one single-shot film, portrait by default. There is no structured project state; write any setting that must be reused across turns into the reply.
## When to use
- The user uploads/points to a reference image and wants a "floating candle" / "jack-o'-lantern" themed Halloween short.
- The user asks for another take, a stronger or weaker fog / light / horror feel, landscape instead of portrait, or a longer story.
- The user sends only an image without saying what they want, or only wants a still image → confirm intent first; do not generate a video by default.
## Steps
1. **Confirm the input**: there must be one reference image (the `asset_id` returned for the user's attachment, or a usable https / `s3://`). If there is no image, ask the user for one first; do not invent one. Default output conventions: portrait `orientation='portrait'`, short clip, no subtitles, no dialogue, pure visual storytelling, delivered as a "narrative preview" rather than a full storyboard reconstruction; only switch to `landscape` if the user explicitly asks for it.
2. **Read the reference image** (recommended): call `recognize_image(image_url=<reference image>, prompt="Describe in English the main subject (appearance, hair, clothing, pose), the environment, the color palette and the lighting, and which elements should be kept as anchors in a video")`. Write down the returned subject description, palette and lighting and put them into the prompt to keep the character consistent. If recognition fails, skip it and go by the user's text description.
3. **Write the video prompt (in English)**: apply the "Look and style conventions" and "Prompt templates" below, replacing `<SUBJECT>` in the template with the subject / outfit / palette from the reference image, keeping the other themed elements fixed and not rewriting the style words. The prompt is in English; no dialogue / subtitles by default (the user's language is used only in the reply text).
4. **Generate**: call
`generate_video(prompt=<the English prompt from the previous step>, input_urls=[<reference image>], model='ref2v', orientation='portrait', name='Floating Candle', entity_type='final_videos')`.
- `ref2v` = reference-to-video: the reference image is the **subject and environment anchor**; it is not required to be the first frame (keeps the character / scene consistent).
- Only when the user explicitly says "make this image itself move, use it as the first frame", switch to `model='i2v'` with the other parameters unchanged.
- `input_urls` needs at least 1 image; do not invent URLs. Do not pass `duration` / `resolution` unless required (use the defaults). Ambient sound / voice-over must be passed as `audio_url`, not stuffed into `input_urls`.
- Video generation does not support NSFW / adult content, with no exception for any channel.
5. **(Optional) Extend the narrative**: when one short clip cannot hold all six beats of "leading the way → knocking → a table of glowing candy", generate 2–3 short videos from the same reference image and the **same style anchor text** (one beat per clip, `entity_type='final_videos'`, same `orientation`), then use
`mix_clips(clips=[{video: <clip 1 asset_id>}, {video: <clip 2 asset_id>}, {video: <clip 3 or audio track>}], orientation='portrait', name='Floating Candle - Narrative Cut', entity_type='final_videos')`
to join them into one film. Clip references may only use the `asset_id`s of assets generated in this conversation (also `s3://` / https); never invent references.
6. **Reply**: summarize the finished film in one or two natural-language sentences (mood, shot, duration, which beats); do not paste long URLs; the film is added to the canvas automatically. When done, call `set_guidance` in the same turn with 2–3 next steps related to this clip (for example "generate another take: thic