pov-cinematic-video · publiczny skill Open Skill

pov-cinematic-video

第一人称 POV 电影视频生成 Skill 将你的剧情、对白和人物设定，转化为电影级第一人称 POV 视频提示词。 该技能严格锁定主观视角、真实摄影运动、人物与场景连续性、动作物理规律、对白口型同步和环境声音，让观众真正代入主角视角。适用于仙侠短剧、沉浸式旅拍、动作剧情、情感片段和第一人称 Vlog 等视频创作。

@bayanaf315

Film

7 1 6

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3927363/2026-09-20/kaq5xhhv.png

# First-Person POV Cinematic Video

## When to use
- The user wants a first-person POV (camera = protagonist's eyes) cinematic video prompt, or the actual video, from a plot / characters / dialogue / narration / reference images / scenes / actions.
- The user wants to revise, regenerate, or produce the next shot continuing from a previous tail frame.

## Reference files
- `references/pov-rules.md`: the complete rule set (highest-priority rules, camera rules, body-appearance rules, interaction, film parameters, composition, performance, physics, continuity, first/last frame, dialogue lock, narration/sound, time progression, input handling, output format, default negative words). **Read it before writing any prompt.**

## Steps
1. Read `references/pov-rules.md` and follow it strictly.
2. If the user provided reference images (character / scene / prop), call `recognize_image` on each to lock face, clothing, hairstyle, props, and environment as hard constraints (rule 14).
3. Integrate all user input (plot, relationships, dialogue, narration, refs, duration, aspect, style, action, sound) without altering core content; never invent dialogue the user didn't provide.
4. Produce **one** complete, copy-paste-ready video prompt following the output format in the reference file (style → camera → scene/time → character/action → camera motion → dialogue/narration → ambient sound → light/camera params → continuity → negative words, ending with the mandatory closing sentence).
5. Optionally render the actual video if the user asks:
   - Reference image(s) provided → `generate_video(model='ref2v', input_urls=[refs], orientation='landscape', prompt=<the prompt>)`.
   - No image → `generate_video(model='t2v', input_urls=[], orientation='landscape', prompt=<the prompt>)`.
   - Continuing from a previous tail frame → `generate_video(model='fl2v', input_urls=[prev_tail, next_head], ...)`.
   - Use `name` = shot title, `entity_type` = `final_videos`.
6. Reply with the full prompt text (and the rendered video if generated). Do not output tables, storyboards, or explanations.

## Notes
- Never modify, polish, shorten, or add to user-provided dialogue; keep it verbatim.
- Strict first-person POV: protagonist never shows face/head/shoulders/torso/back/reflection; hands/forearms/sleeves only appear briefly when the plot requires.
- No BGM, no subtitles, no watermarks, no on-screen text; narration and dialogue must not overlap.
- Keep characters, clothing, props, scene, light, time, action, and spatial position continuous.
- If the user asks for content the platform cannot render (e.g. NSFW/adult video), decline politely and still deliver the prompt text only.
- Report failures honestly; never fabricate URLs.