Pumpkin Lantern Awakening · skill pubblica Open Skill

Pumpkin Lantern Awakening

Pumpkin Lantern Awakening

@GG_Bone

TryOn

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4150474/2026-10-15/q9fmpjcx.jpg

# Pumpkin Lantern Awakening

Turn **one** user-uploaded reference image into a 5-second Halloween fantasy short: a roadside jack-o'-lantern suddenly blazes alight, the character locks eyes with it, green-purple magical flames burst out, magic particles spiral around, and the character smirks as they say "I knew it all along" — one complete "awakening" moment.

This skill has no structured project state: the input is one reference image, the output is one `final_videos` clip, and artifacts are added to the canvas automatically.

## When to use

- The user uploads an image and says things like "make this the Pumpkin Lantern Awakening", "light up the jack-o'-lantern in this image", "a Halloween short video", "image-to-video from this reference".
- The user wants to redo/tweak an existing clip of this kind: change camera angle, change lighting, change the line, add or remove props, fine-tune duration.
- The user only sent an image without details, but the topic is clearly this Halloween "awakening" short.

## Input and deliverable

| | Convention |
|---|---|
| Input | **One** user-uploaded reference image (a character/person image works best); may come from this turn's attachment `asset_id`, an existing conversation asset, `s3://`, or https |
| Deliverable | One **5-second short video** (single shot), `entity_type='final_videos'` |
| Generation | `generate_video` + `model='ref2v'` (reference-to-video, preserves character consistency) |

If there is no reference image: explicitly ask the user to upload one (e.g. "Send me a character/person image and I'll make them lock eyes with the jack-o'-lantern"), **do not** invent a character from text alone, and do not fabricate an image link and pretend one was received.

## Steps

1. Get the reference image (usually the one the user uploaded this turn). Confirm you use only **one** main reference image as the character basis.
2. Using the "Video prompt template" below, turn the user's request into an **English** prompt: only replace the `[...]` scene/camera details in the template, and **never restate the character's appearance** — the reference image decides that.
3. Call `generate_video` once:

```
generate_video(
  prompt=<full English scene description, see template below>,
  input_urls=[reference image],  # only this 1 image; asset_id / s3:// / https all work
  model='ref2v',                 # reference-to-video channel; do not write a physical model name, and do not pass gen_mode
  duration=5,                    # 5-second clip; automatically downgraded if the platform cap is lower
  orientation='portrait',        # vertical 9:16 for short video; switch to 'landscape' if the reference is clearly horizontal or the user explicitly asks for landscape
  name='Pumpkin Lantern Awakening',  # canvas title; may append the character name
  entity_type='final_videos'     # final clip (not the default skill_generation)
)
```

   - When resolution does not matter, do **not** pass `resolution` (the default `720P` applies); pass `'1080P'` only if the user wants sharper output.
   - Out-of-range parameters (duration/resolution) are clamped by the platform and reported in the returned `param_adjustments`; never hardcode an over-long duration or a batch size.
4. Wait for this call to return: generation usually takes tens of seconds to a few minutes — **do not** poll and do not resubmit the same task.
5. Summarize the clip in one or two natural-language sentences (jack-o'-lantern lights up → green-purple magical flames and magic particles → the character smirks and says "I knew it all along"); **do not** paste long URLs (the artifact is already on the canvas).
6. If you want to suggest next steps, call `set_guidance` in the **same turn** with 2–3 options related to this clip (e.g. "Try another camera angle", "Make the magical flames stronger", "Change the line to something else"), and call no other tools in that turn.

## Video prompt template (English, beats hardcoded)

It must cover all four blocks —