agentsclimarketplace

Video understanding

Skill worldwonderer/video-recap-skills/skills/video-understanding

把视频分析为结构化理解索引:场景检测、ASR 转写、逐场景 VLM 观察、静音窗口、融合时间线和写作 brief。 用于理解、索引或总结视频,也作为后续创作前的分析阶段。输入视频文件;输出 scenes.json、 asr_result.json、vlm_analysis.json、silence_periods.json、timeline_fusion.json、agent_narration_brief.md。 触发词:视频理解、视频分析、视频索引、video understanding、analyze video、看懂视频。From its SKILL.md

Install
npx -y skills add worldwonderer/video-recap-skills --skill video-understanding

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • reads credentialsReads from 1 credential source: `MIMO_API_KEY`.
  • runs commandsInstructs the agent to run 2 commands, including `python3 scripts/understand.py <video> --work-dir <work_dir> [--context "节目名/角色名"] [--scene-threshold 0.1] [--skip-asr] [--mimo-video-overview] [--force]` and 1 more.

SKILL.md

3.8 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it

1. 定位

本技能把源视频转成 Agent 与下游阶段可读取的理解索引。它的创作角色是素材观察员 / 场记,不是导演:

  • 先观察,再解释;事实与推断分开。
  • 除了“发生了什么”,还要让下游看见知识、权力、目标、关系或情绪在哪一刻变化。
  • 标出由谁的 POV 承载变化、哪个反应或表演不可替代,以及哪里存在完整台词/动作的自然剪辑边界。
  • 证据不足时保留不确定性,不制造戏剧结论。

2. 处理阶段

  1. 场景检测:写 scenes.json,包含切点、时长和废片段过滤结果。
  2. 抽帧:为视觉分析提取代表帧。
  3. ASR:通过 mimo-v2.5-asr 写时间戳对白 asr_result.json
  4. 静音检测:写 silence_periods.json,标注安静窗口与 has_speech
  5. VLM 观察:写 vlm_analysis.json,包含场景描述、深层分析和 frame_facts
  6. 时间线融合与创作 brief:写 timeline_fusion.jsonasr_writing_chunks.jsonagent_narration_brief.md

各阶段只有在输出产物与 provenance sidecar 同时匹配当前视频及影响结果的设置时才会复用;--force 强制重算。

3. 环境要求

# ffmpeg: brew install ffmpeg | apt install ffmpeg | choco install ffmpeg
export MIMO_API_KEY=***

ASR 使用 mimo-v2.5-asr;VLM 使用 mimo-v2.5--skip-asr 可跳过对白转写,但完整理解仍需要 MIMO_API_KEY 运行 VLM。--mimo-video-overview 可开启按场景块的视频概览。

work_dir/background_research.json 存在,本技能会把剧情梗概和角色名折入 VLM 上下文;--context 可补充一条简短提示。

下面的 scripts/... 均相对于本技能目录。若执行器从仓库根目录启动,请给脚本路径加上本技能的绝对目录。脚本不从其他技能目录读取文件;外部输入仅限命令显式传入的视频、参数与 work_dir 产物。

4. 运行命令

python3 scripts/understand.py <video> --work-dir <work_dir> \
  [--context "节目名/角色名"] [--scene-threshold 0.1] [--skip-asr] [--mimo-video-overview] [--force]

5. 输出契约

文件内容
scenes.json场景切点、起止时间与时长
asr_result.json[{start, end, text}] 时间戳对白
vlm_analysis.json逐场景描述、深层分析与 frame_facts
silence_periods.json[{start, end, duration, has_speech}] 安静窗口
timeline_fusion.jsonVLM、ASR 与静音信息的统一时间线
asr_writing_chunks.json按句界和场景切分的 ASR 写作块
agent_narration_brief.mdAgent 首先阅读的创作简报

后续写作阶段根据创作简报与索引制定方案并写 narration.json

6. 参考资料

  • 背景调研:references/research-guide.md,产出 background_research.json
  • JSON 结构:references/data-schema.md

7. 能力边界

  • 不写解说词,也不做解说评分;只负责生成理解索引与创作简报。
  • 不剪辑、不配音、不合成视频。
  • 不编造信号无法支持的剧情;当 ASR / VLM 过薄时输出素材警告。
  • 不发布、不调度,只向 work_dir 写产物并停止。

What ships with it: 25 files

323.4 KB alongside SKILL.md, 22 of them executable

scripts/

Gives 1 of the 12 instructions most research analysis skills give in ~1.2k tokens

Counted across 1,213 of the 2,113 authors here whose files we hold, read 2026-09-06

  • Cite sources for every important claimin 47 of 1213, across 38 files
  • Separate facts from inferences and recommendationshere, and in 21 of 1213, across 12 files
  • Write findings to a markdown filein 19 of 1213
  • Label every insight with a confidence levelin 18 of 1213, across 8 files
  • Read product marketing context before asking questionsin 18 of 1213, across 8 files
  • Rank themes by frequency and intensityin 16 of 1213, across 6 files
  • Establish research mode before proceedingin 16 of 1213, across 6 files
  • Segment survey responses by customer tier or tenurein 16 of 1213, across 6 files
  • Categorize support tickets before analyzingin 16 of 1213, across 6 files
  • Weight research sources from the last twelve monthsin 16 of 1213, across 6 files
  • Use at least five data points per segmentin 15 of 1213, across 5 files
  • Extract verbatim quotes for all research findingsin 15 of 1213, across 5 files

Said here and by no other author read

  • identify POV changes and key reactions
  • mark natural editing boundaries
  • preserve uncertainty when evidence is insufficient
  • use background research for VLM context

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.