Epic Water Slide Video · Open Skill 公开技能

Epic Water Slide Video

Epic Water Slide Video 是 Digen 上的公开 Open Skill。打开后即可了解该工作流能制作什么。

@eggs

Doc

0 1 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/dev3tz4t.jpg

# Epic Water Slide Video

Turns a user's reference image (plus optional text) into a short, high-energy "epic water slide" video: the subject from the picture rides a gigantic, cinematic water slide.

There is no structured project state. Everything lives in the conversation plus the canvas assets produced by the tools; reuse earlier assets by passing their `asset_id` or `s3://` uri as tool inputs.

## When to use

- The user uploads a reference image and asks for a short video of it (a water slide, a water park, "the epic water slide video", "make this person slide down").
- The user uploads an image plus text describing the ride they want (speed, mood, ending, angle) and wants it as a video.
- The user only wants to change one thing about a clip that already exists (the keyframe look, the camera move, the splash at the end, the length feel) — in that case only redo the matching step.
- Text-only requests are supported as a fallback: build the base keyframe with text-to-image first, then animate it.

## Steps

1. Read the guides before doing anything else, with `read_skill_file`:
   - Always read `references/style-and-prompt.md` (style, subject, layout, must-keep/must-avoid, prompt fragments) and `references/workflow-stages.md` (stage split, sub-steps, checks, retry/branch rules).
   - Read `references/video-content-constraints.md` right before every `generate_video` call.
   - Read only what the current stage needs — do not open all reference files at once, and do not re-read a file you already read in this conversation.

2. If the task is multi-step (keyframe + animation, or a revision that touches both), start with `write_todos` listing the stages, and update the whole list as you advance (at most one `in_progress`).

3. Stage 2 — build the keyframe (see `references/workflow-stages.md` for sub-steps and checks):
   - With a user reference image: `generate_image(prompt=<image prompt assembled from references/style-and-prompt.md>, input_urls=[reference image], model='i2i.hd.gpt', orientation='portrait', name='Water Slide Keyframe', entity_type='storyboard_images')`.
   - Without a reference image (text-only fallback): same call but `model='t2i.hd.gpt'` and no `input_urls`.
   - Keep the subject's identity and recognizability intact — the prompt must be built from the style guide's fragments, not improvised.
   - Show the result to the user in words; if the keyframe clearly breaks must-keep rules, re-run this step once with a corrected prompt before animating.

4. Stage 3 — animate it into the short video:
   - `generate_video(prompt=<video prompt assembled from references/style-and-prompt.md and references/video-content-constraints.md>, input_urls=[keyframe asset, original reference image], model='ref2v', orientation='portrait', duration=5, name='Epic Water Slide', entity_type='final_videos')`.
   - `ref2v` needs at least 1 reference image; put the freshly generated keyframe first so the scene and the user's subject stay consistent.
   - Use `orientation='portrait'` for short-video delivery by default; switch to `landscape` only if the user explicitly asks for a horizontal clip. Leave `resolution` unset unless the user asks for higher quality.
   - Only when the user explicitly asks for a fixed start and end frame, use first/last-frame mode instead: `generate_video(prompt=..., input_urls=[start frame, end frame], model='i2v', orientation='portrait', duration=5, name='Epic Water Slide', entity_type='final_videos')`.
   - Do not pass a physical model name anywhere; channels only.

5. Reply in one or two natural sentences describing the clip (subject, ride, mood, ending) — never paste long URLs or raw asset ids. If the clip is done and a next action makes sense (another take, a longer version, a different angle), call `set_guidance` in the same turn as the final text and call no other tool in that turn.

## Notes

- Video generation does not support NSFW/adult content under any channel. Politely decline such requests without ca