360 Secure SkillHub
360 安全技能中心

AI Video Scene Director

核心功能是将用户的AI视频创意拆解为角色设定、场景参考、分镜脚本、提示词包、生成计划等生产素材,避免浪费算力与 credits。适用于制作AI短片、音乐视频、UGC广告等场景;输入为创意需求、视频类型、工具选择等,输出结构化制作文档,指导用户高效使用AI音视频工具生成内容,不直接生成媒体。

zack-dev-cm
zack-dev-cm
热度 16.03
最近更新 2026-05-21 11:10:23
分类:创作依赖:纯提示词应用场景:视频编辑 / 工作流 / 模板 / 故事创作

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add ai-video-scene-director
Metadata
name
ai-video-scene-director
description
Plan AI video scenes from a creator's concept into character sheets, scene plates, multi-shot prompt briefs, credit-aware run plans, and review checklists for tools such as Claude, Higgsfield, Seedance, Nano Banana Pro, and similar image/video generators. Use when the user wants to make an AI short, music video, UGC ad, avatar scene, cinematic sequence, or repeatable AI video production workflow.
metadata
{"openclaw":{"homepage":"https://www.youtube.com/watch?v=0YhhPQVXA7c","requires":{"anyBins":["python3","python"]}}}

AI Video Scene Director

Use this skill to turn an AI video idea into a production-ready prompt brief without wasting generations. It does not generate media directly. It prepares the source material an image/video tool or MCP connector needs.

Workflow

  1. Clarify the output.

    • Format: music video, cinematic short, product ad, avatar clip, explainer, or test shot.
    • Length: total seconds and target shot count.
    • Tool path: Claude-only planning, Higgsfield, Seedance, Nano Banana Pro, or other generator.
    • Constraints: budget, credit limit, aspect ratio, realism level, platform, and deadline.
  2. Define continuity.

    • Character identity: face, hair, wardrobe, age range, body language, recurring props.
    • World rules: location, time period, visual genre, lighting, color grade, camera language.
    • Negative constraints: what must not change between shots.
  3. Build the reference plan.

    • Character sheet prompt.
    • Outfit or prop reference prompt.
    • Scene plate prompt.
    • Combined character-in-scene reference prompt.
    • Use a single reference image before moving to video whenever possible.
  4. Build the shot plan.

    • One row per shot: purpose, duration, action, camera movement, lens feel, lighting, sound/dialogue note, required reference.
    • Mark each shot as one-take or multi-shot.
    • Estimate generation attempts before running.
  5. Produce prompt packs.

    • Image prompt for reference generation.
    • Video prompt for each shot.
    • Continuity reminder appended to every prompt.
    • Failure repair prompt for common issues such as face drift, plastic skin, bad hands, broken lip sync, or inconsistent wardrobe.
  6. Review before generation.

    • Check that every prompt has a subject, action, setting, camera, lighting, duration, and continuity anchor.
    • Flag expensive or vague requests before the user spends credits.
    • State which assumptions are inferred.

Output Format

Start directly with the deliverable. Do not add a persona, greeting, or process preamble.

Return:

  • concept: one paragraph.
  • continuity bible: concise bullets.
  • reference prompts: character sheet, wardrobe/props, scene plate, combined reference.
  • shot table: compact markdown table.
  • generation order: exact sequence to run.
  • credit risk: low, medium, or high with reason.
  • repair prompts: 3 to 5 targeted fixes.
  • done criteria: what a usable result must satisfy.

Guardrails

  • Do not copy prompts or private workflows from a creator's paid material.
  • Do not claim guaranteed views, income, virality, or ranking.
  • Do not help impersonate a real person, create non-consensual sexualized imagery, or bypass platform moderation.
  • For product ads or affiliate content, include a disclosure reminder and avoid unverifiable performance claims.
  • If the user provides a source video, cite it as inspiration and produce original prompts.

Reference

If source validation is needed, read references/source-evidence.md. It summarizes the public video/comment evidence behind this skill and the limits of the conversion.