Carve Your Jack-o'-Lantern · skill Open Skill công khai

Carve Your Jack-o'-Lantern

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

@Promise

Custom

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-10-14/fqvx59f3.jpg

# Carve Your Jack-o'-Lantern

Turn the **main subject** from the user's uploaded photo into a **carved cut-out pattern** on a realistic pumpkin surface, lit like a classic jack-o'-lantern. Output is a single hero image on the canvas unless the user asks for variants.

There is no structured project state—reuse the same uploaded image URL or asset id across turns.

## When to use

- User uploads (or already has in chat) a photo and wants a **pumpkin lantern / jack-o'-lantern** with that subject **carved into the pumpkin**.
- User asks for Halloween pumpkin art, "carve my face/pet/logo on a pumpkin," or similar.
- User wants **another version** (scarier, cuter, different angle, different carving depth) using the same reference photo.

## Prerequisites

- At least **one reference image** with a clear **subject** (person, pet, object, logo, character). Prefer a simple background and a recognizable silhouette.
- If the user did not upload an image, ask for one before calling `generate_image`.

## Model selection (image-to-image only)

Always use **`generate_image` with `input_urls`** (the user's photo first). Set **`model`** to one of these channels—**do not** use physical model names or `t2i` for the main carve:

| Channel | When to use |
|---------|-------------|
| `i2i.hd` | **Default** for the final carved pumpkin—best detail in the cut lines and pumpkin skin. |
| `i2i.hd.lite` | User wants a **quick preview**, draft, or lower cost; or first attempt before an HD pass. |
| `i2i.hd.gpt` | Subject is **complex** (busy scene, multiple faces, intricate logo) or a prior `i2i.hd` / `i2i.hd.lite` result lost subject likeness; also when user explicitly asks for the "GPT" / smart edit tier. |

If unsure after a failed or off-subject result, retry once with the **next tier up** (`lite` → `hd` → `i2i.hd.gpt`). Do not call `list_models` unless all three channels fail or the platform rejects the channel name.

All three `.hd*` image channels **do not support NSFW**—refuse adult or explicit requests.

## Workflow

1. **Confirm intent** in one short message: subject from their photo → carved pumpkin lantern look. Note optional preferences (cute vs spooky, indoor/outdoor, square vs portrait).
2. **`write_todos`** when generating more than one variant or running preview then final.
3. **Understand the subject (recommended):** `recognize_image(image_url=<user upload>, prompt="Describe only the main subject, pose, and silhouette suitable for a pumpkin stencil. Ignore background.")` Use the description to tighten the carve prompt; do not paste a long analysis to the user.
4. **Generate the carved pumpkin:** `generate_image` with:
   - `input_urls`: `[<user image asset_id, https, or s3://>]` — first image is the identity/subject reference.
   - `model`: per table above (default `i2i.hd`).
   - `prompt`: English; see **Prompt recipe** below. Fold in `recognize_image` details (species, pose, key edges).
   - `name`: short label, e.g. `Jack-o'-lantern – [subject]`.
   - `entity_type`: `skill_generation` (or `props` if the canvas groups Halloween assets).
   - `orientation`: `square` for classic product-style pumpkin shots unless the user asked for `portrait` / `landscape`.
   - Omit `resolution` unless the user asked for higher than default (then `1080P` or `2K` as appropriate).

5. **Reply:** Briefly describe what was carved and invite one refinement (lighting, expression, more/less scary). Do not dump raw URLs.

6. **Optional second pass:** If the user wants higher fidelity after a `i2i.hd.lite` preview, regenerate with `i2i.hd` or `i2i.hd.gpt` using the **same** `input_urls`, updated prompt from their feedback.

For extra prompt patterns and negative cues, `read_skill_file(carve-your-jack-o-lantern, references/prompt-recipe.md)`.

## Prompt recipe (core)

Build one cohesive English prompt. Always include:

- Real **orange pumpkin** with stem, natural skin texture, photographed realistically.
- The **subject rendered as carved holes** (cut-out stenci