Halloween Contrast Costume Video · সর্বজনীন Open Skill
Halloween Contrast Costume Video
Halloween Contrast Costume Video Digen-এ একটি সর্বজনীন Open Skill। খুলে দেখুন এই ওয়ার্কফ্লো কী তৈরি করে।
Roleplay
1 0 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-19/ft2xpwg9.jpg
# Halloween Contrast Costume Video Turn one uploaded reference photo into a short vertical video: the subject from the photo keeps their identity and framing but is restyled into a high-contrast Halloween costume, played out with a slow reveal and a subtle Halloween camera move. There is no structured project state — the uploaded image, the generated still and the generated video are the only artifacts, and they live on the canvas. ## When to use - The user uploads a photo and asks for a "Halloween transformation", "Halloween costume video", "scary version of this photo", "costume reveal video". - The user only wants the transformed Halloween still image (Stage 2 only). - The user wants to redo one element of a previous result (costume, mood, lighting, camera move) without redoing everything. ## Steps 1. **Confirm inputs.** One clear reference photo is required (portrait or full-body, single main subject). If the user sent several, ask which one to transform. If no photo is available, ask for one — do not invent a subject. 2. **Read the style contract before any generation.** Call `read_skill_file` with `path="references/style-and-prompt.md"`. Every prompt you build (image or video) must follow the Style / Subject / Layout / Must keep / Must avoid rules and the templates in that file. 3. **Analyze the reference (one cheap call).** Call `recognize_image(image_url=<uploaded asset>, prompt="Describe the main subject: face features, hairstyle, clothing, accessories, pose, body framing, background, props, lighting and color grade.")`. Use the answer to fill the "original look" slots of the prompt templates. Skip only if the user already described the photo in detail. 4. **Stage 1 — Halloween costume still (image-to-image).** Call: `generate_image(prompt=<image-to-image template from references/style-and-prompt.md, filled with the analyzed look>, input_urls=[<uploaded asset>], model='i2i.hd.lite', orientation='portrait', resolution='1080P', name='Halloween costume still', entity_type='storyboard_images')` The still must keep the person's identity and framing while pushing the costume/mood contrast. Check the result against the Must keep / Must avoid list; if a hard constraint is broken, retry once with the failing clause made more explicit before moving on. 5. **Read the motion contract.** Call `read_skill_file` with `path="references/video-content-constraints.md"` and follow its shot, reveal beat, camera move, pacing and forbidden-content rules. 6. **Stage 2 — short video (reference-to-video).** Call: `generate_video(prompt=<video template from references/video-content-constraints.md>, input_urls=[<Halloween still from step 4>], model='ref2v', orientation='portrait', resolution='1080P', duration=5, name='Halloween contrast reveal', entity_type='final_videos')` Pass the Halloween still as the reference image so identity and costume stay consistent. If the user explicitly wants the untouched photo carried through instead, pass the original upload as the reference as well (still first, photo second). 7. **Delivery.** Describe the result in natural language: costume direction, mood, camera move, and one or two concrete variants the user could ask for. Do not paste long URLs. 8. **Progress and guidance.** For the full pipeline call `write_todos` up front (analyze / still / video / deliver) and update the whole list as stages complete. At a stage boundary or when everything is finished, call `set_guidance` in the same turn as the final text with 2–4 relevant next actions (e.g. "Make the costume scarier", "Try a witch variant", "Change the reveal move"). ## Notes - The reference photo is the subject of the whole app: never replace the person, never swap the face for a different character. Halloween styling is applied **on** that subject. - Costume direction comes from the reference (clothing, accessories, palette, setting) plus the user's wish. If the user names a costume, honour it; otherwise pick the variant in