Falling Witch Hat · публичный навык Open Skill
Falling Witch Hat
Falling Witch Hat
TryOn
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-10-15/jys8kere.jpg
# Falling Witch Hat Turn the user's **single reference image** into a short video: an autumn night, leaves drifting, a witch hat wrapped in a violet magic swirl descending from the sky and landing precisely on the character's head — the coronation moment of the "chosen one" — ending with the character giving a mischievous grin and raising a hand to straighten the hat. The whole clip serves exactly one theme: **frame content, palette, materials, and lighting are all unified around the "falling witch hat"**, and it must not expand into any other story. There is no structured project state; the generated video lands on the canvas, so the final reply can just summarize in natural language. ## When to use - The user uploads/provides an image and says "make it into the falling witch hat", "drop the witch hat onto his head", "crown him", etc. - The user wants to redo or fine-tune an existing version: swap the reference image, switch orientation, change the leaves/background, change the hat color or material, make the motion slower. - Not applicable: pure text-to-video (no reference image), static images only, or editing multiple existing clips — this skill only does a **single-shot reference-to-video**; for those requests, state the capability boundary directly. ## Input - **At least 1 reference image** is required (an image asset uploaded in this turn; `asset_id` / `s3://` / https are all accepted). If the user only gave text and no image, ask them to upload one first — **do not** invent the subject's appearance. - If multiple images are given: pick the one with the clearest subject, largest figure, and no occlusion as the main reference (put it in `input_urls[0]`) and ignore the rest. - If the reference image contains nudity or other adult content: politely refuse and explain that video generation does not support NSFW/adult content; do not call the tool. ## Steps 1. **Read the image and pin down the subject**: confirm the character's appearance (face shape, hairstyle and hair color, eye color), outfit, pose, and the existing art style, then pick 1–2 features that **must be preserved** (for example "silver-white long hair + deep blue cloak") to write into the prompt. 2. **Write one English prompt following the "Prompt template" below**, filling in all six sections one by one — do not skip sections and do not write just a single sentence. 3. **Call the generation tool**: ``` generate_video( prompt = <the English prompt from the previous step>, input_urls = [<reference image>], # at least 1; the first is the main reference model = 'ref2v', # reference-to-video channel: keeps the subject consistent orientation = 'portrait', # portrait by default for short video; only change to 'landscape' if the user explicitly asks name = 'Falling Witch Hat', # or '<subject name> · Falling Witch Hat' entity_type = 'final_videos', ) ``` - Do not pass `gen_mode` (the mode is fully determined by `ref2v`), and do not pass a physical model name. - Do not proactively pass `duration` / `resolution`; use the channel defaults (720P). Only pass them when the user explicitly asks for "slower / longer / higher quality", and never write an over-long duration or an over-high resolution — out-of-range parameters are automatically clamped and reported in the returned `param_adjustments`. - Generation takes a while; do not resubmit the same task before this call returns. 4. **On success**, summarize in one or two natural-language sentences (who, what hat, what atmosphere) and **do not paste long URLs**; in the same turn you may call `set_guidance` with next steps related to this output (for example "swap in another reference image and make another version", "make the hat fly in from the side and redo"). 5. **On failure** (`success=false`), explain the error honestly — do not fabricate a video link and do not pretend it succeeded. ## Prompt template (English, write every section) One-line ma