Communitygithub.com

CPT-Dawn/souvi

使用 SouVi 引擎(`souvi` CLI)创建和编辑视频与音频——涵盖剪辑、短片、预告片、幻灯片、调色、字幕、旁白、音乐、混音、母带处理、音效设计和格式导出。

souvi 是什么?

SouVi Engine(`souvi` CLI)是一个功能全面的命令行视频和音频制作工具集,专为 AI 代理和开发者设计。它支持从导入素材、剪辑、调色、加字幕、配乐/旁白,到混音、母带处理、音效设计以及多种格式导出(如预告片、短片、幻灯片等)的完整工作流。该项目提供了清晰的项目结构(`projects/<slug>/`、`assets/<slug>/`、`output/<slug>/`),确保素材不被修改、渲染版本化且永不覆盖;此外,它通过文件锁支持并行代理安全操作。用户可通过 `souvi new`、`souvi import`、`souvi plan`、`souvi render` 等命令快速上手,并利用 `souvi look`、`souvi hear` 和 `souvi frame` 等工具以连续帧、波形图等方式“查看”或“聆听”媒体内容,从而进行审核。SouVi Engine 可在任何目录下运行(引擎位于 `/home/cptdawn/Kode/SouVi-Engine`),适用于主流 AI 编码环境(如 Claude Code、Codex CLI、OpenAI Codex CLI 等),且鼓励遵循 `video.md` 和 `audio.md` 中的最佳实践以及 QC 检查清单,确保专业级输出。

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/CPT-Dawn/SouVi-Engine/tree/HEAD/skill

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

SouVi Engine

souvi is on PATH and works from any directory. The engine lives at /home/cptdawn/Kode/SouVi-Engine (below: $SOUVI). souvi <cmd> -h gives flags; souvi presets lists delivery formats, looks, title styles, transitions and sfx.

Workspace

Each deliverable set is a project in $SOUVI:

  • projects/<slug>/ — timeline.json (the edit) and BRIEF.md (brief + dated decisions log)
  • assets/<slug>/{footage,images,music,voice,sfx,graphics,fonts,luts,refs}/ — sources, never modified
  • output/<slug>/{drafts,renders,final,review}/ — results, versioned _vNNN, never overwritten

Called from another repo: slug = that repo's folder name (suffix it for a separate deliverable, e.g. myapp-teaser). Import media with souvi import <slug> <files…> (--as voice|sfx for non-music audio, --link for large footage); deliver with souvi promote <render> --to <folder in the calling repo>. Leave git in $SOUVI alone unless the user asks.

Parallel agents are safe: every job runs in a private folder and renders of one project queue on a lock. Give each agent's deliverable its own slug.

Loop

  1. souvi new <slug> --preset <delivery>; write the brief into BRIEF.md.
  2. souvi import …, souvi assets <slug>, then souvi look <clip> on every candidate source.
  3. Write timeline.json after reading $SOUVI/docs/timeline.md; run souvi plan <slug> until clean.
  4. souvi render <slug> --draft (--gpu, --until N), review, revise.
  5. souvi render <slug> (QC must PASS), review, souvi promote <render> [--to …].

Review — you cannot watch or listen, so read the PNGs these print: souvi look <video> (contact sheet; --from/--to for dense sheets around cuts), souvi frame <video> --at <t…>, souvi hear <file> (spectrogram, waveform, loudness). Work is done when the final render passes this review and QC.

Craft

Before creative decisions, read video.md for picture and audio.md for sound. Apply every item of their review checklists.

Single-file tools

export --preset, master, voice, sfx, reverb, stabilize, denoise, grade, retime, qc, loudness, probe. Processed files land beside their source as <name>.<tag>_vNNN.<ext>. For anything else, raw ffmpeg is fine when it keeps 10-bit processing, lanczos scaling, BT.709 tags and soxr resampling; finish loudness with souvi master and encode deliverables with souvi export.

相关技能