Vegetable Jack-o'-Lantern · skill Open Skill công khai

Vegetable Jack-o'-Lantern

Vegetable Jack-o'-Lantern là một Open Skill công khai trên Digen. Mở ra để xem workflow này tạo gì.

@eggs

TryOn

1 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/fgu6x5qf.jpg

# Vegetable Jack-o'-Lantern

Text in, one image out. The user names a vegetable (or describes a short scene), and this skill produces a single Halloween night image: a roadside jack-o'-lantern carved from that vegetable suddenly lights up, and a character standing on the road looks straight at it. There is no structured project state — everything lives in the user's message, the prompt you build, and the generated canvas asset.

## When to use

- The user names a vegetable/produce and wants it turned into a jack-o'-lantern or a glowing Halloween lantern image.
- The user asks for a Halloween night scene of a carved glowing vegetable lantern and someone looking at it.
- The user wants a variation of that scene (different vegetable, different costume or props, different framing).

## Input and output

- Input: **text only**, and it is normally just the produce name (the app's `"vegetable name"` slot). Sometimes the user writes a short scene description instead, or pastes the app's own sentence with the placeholder left unfilled.
- Output: **one image asset** on the canvas. Never answer with text only and never call a video tool — this app is image-only.

## Steps

1. Extract the vegetable from the user's message.
   - Bare name (e.g. "carrot", "broccoli", "bell pepper", "pumpkin") → that name goes into the `{vegetable}` slot verbatim; if the user wrote it in another language, translate it to the common English produce name.
   - Short description → keep every detail the user named (vegetable, costume, extra props, mood) and fill the rest from the template.
   - The placeholder is left empty or literally says "vegetable name" → ask one short question: which vegetable should the lantern be carved from? Then continue.
2. Before generating, read the visual bible once: `read_skill_file('vegetable-jack-o-lantern', 'references/style-and-prompt.md')`. It holds the Style, content, layout, must-keep and must-avoid rules, the copyable prompt templates and the negative prompt. Read it once per conversation and reuse it; do not reload it on every turn.
3. Build the final English prompt from the template in that file: start from the base template, substitute `{vegetable}` with the user's produce, and paste in any extra detail the user added. Do not shorten the template down to one line, and do not drop the eye-contact beat, the Halloween props or the lighting rule.
4. Call `generate_image` exactly once per requested image:
   - `prompt` = the assembled English prompt
   - `model='t2i'`
   - `orientation='square'`, `resolution='1080P'` by default; switch to `portrait` or `landscape` only if the user explicitly asks for a phone wallpaper / horizontal shot
   - `name` = short canvas title, e.g. `Vegetable Jack-o'-Lantern — Broccoli`
   - `entity_type='props'`
5. Reply in one or two natural sentences: name the vegetable and confirm the lantern is lit and the character is looking at it. Do not paste URLs or the raw prompt.

## Notes

- Keep the named vegetable recognizable. The joke of this app is a *vegetable* carved into a jack-o'-lantern — never silently swap it for a pumpkin, and never blur it into an anonymous gourd.
- Keep the Halloween theme everywhere: props, decor and light source all belong to Halloween night. No daytime, no summer garden, no neutral studio lighting.
- Friendly spooky only. Refuse gore, real blood, sexual or otherwise unsafe content and do not call the tool; video generation is not available for this app anyway.
- If `generate_image` returns `success=false`, say so plainly and do not fabricate an image or a link. One call produces one image, so make a separate call per variant instead of asking for several in one prompt.
- The user may change one thing (the vegetable, the costume, the framing). Re-read the reference file only if it is no longer in context, then regenerate with that single change applied and everything else preserved.