360 Secure SkillHub
360 安全技能中心

MovieStar

MovieStar 是确定性视频编辑层,通过 CLI 管理项目状态、时间映射、渲染与结构化验证。适用于修剪、转录搜索、多机位合成、场景布局、字幕、叠加、混音和导出检查。输入源视频与编辑指令,输出媒体文件、JSON 状态和验证结果。源文件不被修改;需在任务目录操作,并建议先 doctor/probe,采用检查-规划-应用-验证循环,适合需可复现剪辑和验证的用户。

jdilla1277
jdilla1277
热度 7.62
最近更新 2026-09-28 17:31:02
分类:创作依赖:代码执行 / 读写文件应用场景:视频编辑 / 工作流

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 moviestar
Metadata
name
moviestar
description
Edit and verify videos with the MovieStar CLI. Use for trimming, transcript search, multicamera composition, scene layout and motion, captions, overlays, audio mixing, rendering, or inspecting MovieStar projects and exports. Do not use for image-only editing or unrelated FFmpeg administration.
metadata
{"openclaw":{"requires":{"bins":["moviestar","ffmpeg","ffprobe"]},"homepage":"https://trymoviestar.com"}}

MovieStar

Use MovieStar as the deterministic editing layer. Make editorial decisions from the user's request and inspected media; let the CLI own project state, time-mapping, rendering, and structured verification data.

Start safely

  1. Work in a dedicated task directory. MovieStar creates a moviestar/ project there and stores absolute pointers to source media; it never copies or modifies the source files.
  2. Run moviestar doctor. If the command is missing or the environment is not ready, follow the bundled installation guidance.
  3. Run moviestar probe VIDEO when source dimensions, codecs, duration, or audio presence matter.
  4. Use moviestar --help and moviestar COMMAND --help before unfamiliar operations. Do not invent flags from memory.

Edit in an inspect-plan-apply-verify loop

  • Load media once with moviestar load. Name sources with --as when stable IDs will help a multicamera or multi-file edit.
  • Inspect before cutting: use status, find, skim, storyboard, inspect, screenshot, or watch according to the evidence needed.
  • Treat stdout as structured JSON. Read paths and ranges from response fields; progress narration is on stderr. Use --quiet only when a harness combines both streams and requires pure JSON.
  • Preview expensive or replacing operations with --dry-run. Review resolved ranges, durations, warnings, output paths, and FFmpeg plans before applying.
  • Remember the clocks: trim is result-time, find reports both source and result ranges, and composition ranges are documented per command. Never substitute one time space for another.
  • After each material edit, inspect status or the relevant read command. Verify visuals with returned image/video files rather than Base64 text.

Read editing workflows when choosing between single-source, transcript, multicamera, scene, caption, overlay, or audio commands. Before delivery, follow render verification.

Preserve user control

  • Keep source media untouched and write generated media to the task directory or a user-approved output path.
  • Do not use --force, replace a composition/state file, overwrite an explicit output path, or delete generated artifacts without previewing the effect and confirming it matches the request.
  • Report important warnings instead of hiding them: weak transcript matches, framing coverage, softness, clipping, loudness limits, duration drift, and missing dependencies can change the deliverable.
  • Do not submit feedback, create accounts, subscribe an email address, download large models, or install optional dependencies unless the user requested or approved that action.

Finish with evidence

Return the final output path, probed duration and streams, the editing choices made, and any remaining warning. A successful command is not by itself proof that the edit looks or sounds right; inspect representative frames and motion, and check audio when the deliverable contains it.