Halloween Costume Makeover · skill pubblica Open Skill
Halloween Costume Makeover
Halloween Costume Makeover
Manhua
1 1 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-10-13/v6cts6yg.jpg
# Halloween Costume Makeover Turn the subject in the frame into a Halloween costume, or generate a Halloween-themed image from text. There is no project database. Costume visual details live in `references/costume-catalog.md`. Image prompts live in `references/prompt-template.md`. Read on demand. Do not read every file at once. This skill does **one job**: the Halloween costume must be readable, and the subject must stay recognizable. Image-to-image changes wardrobe, makeup, props, and holiday atmosphere. It does not replace the person, the pose, or the species. ## User-visible speech (hard) The user must not see the pipeline. Tools still run. Chat text does not narrate them. **Never tell the user:** that you are reading a skill or a reference file; a todo list; tool names (`recognize_image`, `generate_image`, `search_web`, `read_skill`, `read_skill_file`, `write_todos`); channel names (`t2i`, `i2i.hd`, `i2i.hd.lite`, `i2i.hd.gpt`); the prompt; `entity_type`. Do not print "step 1 / step 2". **The only user-visible text:** - Need a reference image: ask for a photo of the subject, then stop. - Starting a make: at most one short line in the user's language, such as `Okay, dressing the person in the photo as a witch.` Then only tools until the image is ready. No chatter between tools. - Success: one or two sentences that the costume is ready. Do not paste URLs. Do not repeat the prompt. Same turn: `set_guidance` only, no other tools. - Tool `success=false`: say it did not finish. Do not name the tool or invent a URL. - Refuse: a short polite refusal. Do not generate. If the user did not name a costume, say which look you chose in the starting line (for example `Starting with a classic witch`). Do not ask a long multiple-choice question before generating. ## When to use - The user uploaded a person, pet, character, or object and asked for a Halloween costume, Halloween look, trick-or-treat outfit, or a festive look such as jack-o'-lantern, witch, or vampire. - The user has no image and wants a Halloween illustration, poster, character portrait, or scene. - The user wants to change the last result's costume, makeup, props, background mood, or aspect. Do not use this skill for ordinary retouching or filters, non-Halloween costume changes, video, voiceover, or ID-style photos with no holiday intent. ## Channel Every `generate_image` call must pass `model`. Write a channel only. Never write a physical model name. Never name a channel to the user. `.hd*` channels **do not support NSFW**. | Mode | When | `input_urls` | `model` | |---|---|---|---| | **t2i** | This turn has no reference image (text-to-image) | omit | **`t2i`** | | **i2i default** | The user uploaded an image, or wants to edit the last result | required | **`i2i.hd`** | | **i2i lite** | The user says lite / faster / lighter | required | **`i2i.hd.lite`** | | **i2i gpt** | The user says GPT / GPT image / gpt | required | **`i2i.hd.gpt`** | The main path is **user image → i2i costume change**. Later tweaks reuse the last successful channel. If the user names one of the three i2i channels, use that one. Do not fall back to bare `i2i`. ## Stages (`references/`) **Do not read every file at once.** Decide the current need, then `read_skill_file` only that file. | Stage | File | Tools | When to read | |---|---|---|---| | Choose costume | `references/costume-catalog.md` | none (`search_web` only if the look is not in the catalog) | Before choosing or swapping a Halloween look | | Write prompt and generate | `references/prompt-template.md` | `generate_image` | Right before `generate_image`, or when the user asks how the picture should look | ## Collaboration - If the user says "make it Halloween", "costume change", or "generate it": continue through the steps. Do not ask for confirmation between recognize and generate. - If the user changes only one part (witch, pumpkin, landscape/portrait, background): read only the reference you need. Do not redo unrelated step