Halloween Creative Portrait · публичный навык Open Skill
Halloween Creative Portrait
Halloween Creative Portrait — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.
Custom
0 0 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/trgatjeu.jpg
--- allowed-tools: - generate_image - recognize_image --- # Halloween Creative Portrait Turn one uploaded photo into a single Halloween-themed creative image. The uploaded photo stays the hero: its subject, pose and outfit are preserved, while the world around it is re-dressed with jack-o'-lanterns, fog and amber backlight in a vertical poster composition. This Skill has no structured project state. Everything you need is in the conversation, the uploaded reference image on the canvas, and the reference files below. ## When to use - The user uploads a photo and asks for a Halloween version, Halloween poster, "spooky me", costume/pumpkin themed artwork, or a Halloween short-film style key visual. - The user already got a result from this Skill and asks for a variation: more pumpkins, different costume, warmer/ colder light, different aspect, "make it scarier / cuter / more cinematic". - The user asks what this app does and then hands over a picture. Do not use it for: plain photo retouching with no Halloween theme, text-only poster design, or video output (this app produces a still image only). ## Inputs and outputs - Input: one reference image (the user's uploaded photo). If several images are uploaded, treat the clearly human/subject-bearing one as the source; ask only if it is genuinely ambiguous. - Output: exactly one still image, vertical by default, delivered through `generate_image` and shown on the canvas. - Every input image must be passed to the tool as `input_urls=[<source image>]`. Never describe the source in words only and then generate from text — that loses the subject. ## Steps 1. **Confirm the source image.** Identify which uploaded image is the subject. If there is no image at all, ask for one and stop; do not invent a subject. Read `references/step-details.md` for the intake checklist (multi-person photos, group shots, unclear subjects). 2. **Read the visual rules.** Call `read_skill_file` with `path="references/style-and-prompt.md"`. This file holds Style, Subject/content, Layout, palette, prop conventions, Must keep / Must avoid, and the fill-in English prompt template. You must read it before any generation call and follow it literally. 3. **Inspect the source when needed.** If the photo is busy, low contrast, or you cannot tell what the subject is wearing, call `recognize_image(image_url=<source image>, prompt="Describe the main subject: person count, pose, hairstyle, outfit, accessories, framing, background.")`. Use the answer only to write a more faithful prompt — never to replace the image. 4. **Pick optional variation details.** If the user asked for a specific mood or extra props (scarier, cute, neon, pumpkin overload, witch hat, haunted mansion), open `references/scene-elements.md` and copy the matching snippets into the prompt. Skip this file when the user wants the default look. 5. **Write the image prompt.** Fill the template from `references/style-and-prompt.md` in English: subject clause (from the source photo), Halloween restyle clause, prop/decoration clause, lighting and fog clause, composition clause, and the Must-keep / Must-avoid tail. Keep it concrete and specific to this user's photo; do not paste a generic paragraph. 6. **Check before generating.** Run the pre-flight checklist in `references/step-details.md` (identity preserved, no invented facts, no text or watermarks requested, no adult content). 7. **Generate.** Call: `generate_image(prompt=<the English prompt>, input_urls=[<source image>], model='i2i.hd.lite', orientation='portrait', resolution='1080P', name=<short descriptive title>, entity_type='halloween_portraits')` - `model='i2i.hd.lite'` is fixed: this app is an image-to-image restyle of an uploaded photo. Do not switch to a text-to-image channel while a source image exists. - `orientation='portrait'` is the default for this app (vertical poster look). Use `orientation='square'` or `landscape` only when the user explicitly asks for that shape and wa