{"slug":"video-cut","title":"video-cut","summary":"把长视频按 Agent 选择的原片区间剪成短片。作为两阶段创作流程中的剪辑环节，读取 clip_plan.json 与源视频， 输出 edited_source.mp4；随后 Agent 按输出时间线写 narration.json。支持单视频与多视频（sources manifest）拼剪， 本工具不读取、不映射旁白。 触发词：视频剪辑、剪辑式解说、video cut、clip plan、拼剪。","platform":"Claude","tags":[],"authorName":"LLM Mart","authorSlug":"llm-mart","score":0,"source":"github","price":null,"verified":false,"createdAt":"2026-10-02T16:32:47.519815Z","repo":{"url":"https://github.com/zenstory-ai/oh-story-dsh","stars":433,"forks":74,"license":"MIT","updatedAt":"2026-10-02T11:40:12Z"},"bodyHtml":"
\n

name: video-cut\nuser-invocable: false\ndescription: >\n把长视频按 Agent 选择的原片区间剪成短片。作为两阶段创作流程中的剪辑环节，读取 clip_plan.json 与源视频，\n输出 edited_source.mp4；随后 Agent 按输出时间线写 narration.json。支持单视频与多视频（sources manifest）拼剪，\n本工具不读取、不映射旁白。\n触发词：视频剪辑、剪辑式解说、video cut、clip plan、拼剪。

\n

1. 定位

\n

本技能只执行 Agent 已经做出的剪辑决定：

\n
    \n
  1. 校验并补全 clip_plan.json，写出带 clip_id、原片/输出时间与时长的 clip_plan_validated.json。
    2. \n
  2. 先避开原片硬切附近的闪帧风险，最后把边界吸附到可靠句末/自然停顿；声音完整性拥有最终优先级。
    3. \n
  3. 拼接选定区间，输出 edited_source.mp4。
    4. \n
  4. 到此停止，由 Agent 按真实输出时间线写 narration.json；本工具不读取旁白，也不做原片→输出映射。
    5. \n
\n

相同输入会得到相同输出。edited_source.mp4.meta.json 记录标准化 clips、渲染设置和每个源文件的 size/mtime_ns；三者与当前一致且 edited_source.mp4 存在非空才复用，任一不同即重渲染。只有 sidecar 而没有媒体文件不复用。

\n

2. 输入契约

\n

work_dir/clip_plan.json 可以是数组，也可以是 {\"clips\": [...]}：

\n
{\"start\": 12.0, \"end\": 28.5, \"reason\": \"b02 | turn | power: A→B | POV=女主 | 保留反应 | 入点=问题落下 | 出点=沉默结束\"}\n
\n\n

3. 剪辑意图契约

\n

工具不会替 Agent 做创作选择。写片段前先完成本节的剪辑意图检查，并让每个区间映射到 recap_story_plan.json 的一个 beat。

\n

使用现有自由文本 reason 保存简洁决定：

\n
beat_id | function | change | POV | preferred moment | 入点 reason | 出点 reason\n
\n

不要因为“事件重要”就保留整段；要保留最能让 change 成立的具体表演、反应、动作或揭示。理解与情绪允许时晚进早出，同时保证台词、动作和技术边界完整。

\n

对不能删去的问答、反应或动作兑现，先核源证据，再在同一 clip_plan.json 登记精确区间：

\n
{\n  \"clips\": [{\"start\": 12, \"end\": 18}],\n  \"required_evidence\": {\n    \"nodes\": [\n      {\"id\": \"refusal\", \"source\": \"/media/episode.mp4\", \"start\": 12.25, \"end\": 14.5, \"track\": \"audio\", \"content\": \"对方拒绝请求\"},\n      {\"id\": \"response\", \"source\": \"/media/episode.mp4\", \"start\": 15, \"end\": 17.5, \"track\": \"video\", \"content\": \"听到拒绝后的反应与决定\"}\n    ],\n    \"before\": [[\"refusal\", \"response\"]]\n  }\n}\n
\n

source 使用实际源文件绝对路径，start/end 是原片秒；多源可另填 source_id 消歧。只登记确实需要保留的具体时刻，不将整个 beat 默认锁死。before 只登记本片必需的先后关系；无需约束顺序时写 before: []。

\n

工具在全部画面/句界吸附后检查每个必保时刻至少有一处完整连续保留、来源和先后；音频节点还检查源音轨是否存在。每次结果出现（包括局部片段）都需满足其声明的前提，不能用后面的完整段替开头缺前提的片段过关。结果写入 clip_plan_validated.json.qc.required_evidence；缺段、错序或无效声明会在预检、缓存复用和渲染前阻断，时长放宽选项不会跳过。该结果验证选段保留，实际语义与最终混音仍按审片步骤核对。

\n

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。

\n

4. 运行命令

\n
python3 scripts/cut.py <video> --work-dir <work_dir> \\\n  [--target-duration 10m] [--clip-padding 0] [--allow-overlap]\n
\n

5. 输出契约

\n\n

下游把 edited_source.mp4 当作视频，把 Agent 按输出时间写的 narration.json 当作旁白。

\n

6. 边界与时间线规则

\n\n

需要检查短时间频繁切镜时，先用 ffmpeg scene filter 召回候选时间：

\n
ffmpeg -i input.mp4 -vf \"select='gt(scene,0.35)',showinfo\" -an -f null -\n
\n

0.35 是起始阈值，不是质量判据；大幅运动、闪白和叠化都可能误报。把候选映射回原片 shot 与本次拼接边界后，按上面的来源分类处理，并以正常速度播放决定是否保留。

\n

需要精确到实际帧、检查长区间内部残镜并记录所用计划路径时，使用\nscripts/shot_review.py 或 cut.py --review-shots（有黑边或包装时加 --roi / --shot-roi）；\n详见 references/shot-review.md。

\n

7. 能力边界

\n\n","files":[{"path":"references/shot-review.md","sizeBytes":5549,"isText":true},{"path":"scripts/cut_cli.py","sizeBytes":12385,"isText":true},{"path":"scripts/cut_contract.py","sizeBytes":16364,"isText":true},{"path":"scripts/cut.py","sizeBytes":1131,"isText":true},{"path":"scripts/cut_render.py","sizeBytes":11249,"isText":true},{"path":"scripts/lib.py","sizeBytes":4796,"isText":true},{"path":"scripts/media_geometry.py","sizeBytes":8555,"isText":true},{"path":"scripts/narration_mapping.py","sizeBytes":2674,"isText":true},{"path":"scripts/narrative_selection.py","sizeBytes":8972,"isText":true},{"path":"scripts/sentence_boundaries.py","sizeBytes":22883,"isText":true},{"path":"scripts/shot_review.py","sizeBytes":17440,"isText":true},{"path":"SKILL.md","sizeBytes":7281,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-10-02T16:33:46.467143Z","sha256":"6FB7632C367A0B97CE88A7D32CDC6056C3C1215CB13B54684449C18455ECF403","sizeBytes":40064},"review":null,"source":{"repositoryUrl":"https://github.com/zenstory-ai/oh-story-dsh","path":"packages/knowledge/video-recap/skills/video-cut","license":"MIT","commit":"d734089bae893783d8bfaf09f8e8fa0a99723ef0","subtreeSha":"8313F2178995184BD68103FABB05DCB96702FD480EA06F2594EDBC09C2E3F24E","lastSyncedAt":"2026-10-02T16:32:40.531554Z"},"reviewedAt":"2026-10-02T16:35:17.992179Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/zenstory-ai/oh-story-dsh/tree/main/packages/knowledge/video-recap/skills/video-cut"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install zenstory-ai-oh-story-dsh@llmmart"},{"target":"git","command":"git clone https://github.com/zenstory-ai/oh-story-dsh.git"}]}