| name | kuaidao-live-video-editing |
| description | Use when turning Chinese livestreams, interviews, podcasts, courses, or long-form talking-head videos into story-complete vertical short videos for Kuaidao-style channels. Covers source ingest, StepFun/StepAudio ASR, topic/window discovery, story-spec gates, 5-10 minute story editing, Chinese subtitle cleanup, cover selection, picture-in-picture assets, HyperFrames packaging, versioning, QA, and teammate handoff. Triggers: 快刀, 直播剪辑, 长视频切短视频, 访谈切片, 课程切片, 书籍种草短视频, 加字幕, 做封面, 画中画, HyperFrames, 视频工作流, 短视频工作流. |
Kuaidao Live Video Editing
快刀短视频剪辑工作流 turns a long Chinese source video into a publishable short video with clean subtitles, strong story structure, versioned outputs, and team handoff files.
Use This For
- 直播原片剪成 5-10 分钟短视频
- 访谈、播客、课程、分享、带货讲解的主题切片
- 从长视频里筛选多条候选短视频
- 需要中文字幕、花字、封面、画中画素材、版本管理和 QA 的视频
- 需要多人协作接力的视频工程
Do Not Use This For
- 只需要简单截一刀的任务
- 没有转写也不需要字幕的纯视觉混剪
- 多机位自动切换本身，那应先用专门的 multicam workflow
- 上传发布平台自动化，本 skill 只到交付文件
Core Rules
- 原片不复制、不覆盖、不重编码作为中间源，尽量用 symlink 和代理文件。
- 先转写再判断，不只靠画面抽帧定位内容。
- 开头 10 秒要抓人，随后必须补足冷启动上下文。
- 字幕是清理后的口语字幕，不是原始 ASR 逐字稿。
- 封面、字幕、工程、成片全部版本化，不覆盖用户已看过的版本。
- 自动化机械动作，保留故事、标题、情绪和剪点判断。
- 用户说“牛逼一点 / 高级一点 / 节奏好一点”不算需求；必须追到具体画面、剪点、观众反应和发布平台。
Startup Mode Check
Before doing production work, inspect the project folder and classify the request:
- New edit mode: no usable
project.yaml or cut-plan.md, or the user is starting from a new source video.
- Continue-edit mode:
project.yaml, cut-plan.md, edit-notes.md, or version logs already exist and the user wants to continue the project.
- Iteration mode: the user asks to revise a reviewed draft, such as subtitles, pacing, PIP placement, cover, music, or story structure.
- Single-surface mode: the user only asks for subtitles, cover, QA, ASR, or packaging.
Do not restart the full workflow when the user is clearly iterating. Read existing notes first, then touch only the affected planning surface.
Flow Weight Check
Before applying the full story gates, choose the lightest workflow profile that can safely finish the user's job:
lightweight-talking-head: one main talking-head or口播 video, usually short or already selected. Use for trimming filler, cleaning pacing, subtitles, simple cover, and direct export. Ask only for platform/duration, keep-remove requirements, subtitle style, and cover title when missing. Skip candidate pool and full storyboard unless the user asks.simple-animation: no long source selection; the user wants a short text/data/UI animation, explainer, or voiceover-driven clip. Ask for core message, duration/aspect, voiceover path, and visual style. Create a 3-8 scene storyboard, not a livestream cut plan.standard-clipping: long livestream, interview, course, podcast, or topic extraction from a larger source. Use candidate windows, story arc, subtitle plan, visual plan, and QA.heavy-collab: multi-version, teammate handoff, high-stakes publishing, multiple assets, or unclear source coverage. Use the full gate set, project contract, version log, and review workflow.
Default routing:
- If the user says 口播 / 对镜 / 帮我剪掉废话 / 加字幕 / 出封面, start with
lightweight-talking-head.
- If the user says 简单动画 / 文字动画 / 做个视频说明 / 配音加画面, start with
simple-animation.
- If the user says 长直播 / 访谈切片 / 课程切片 / 从几个小时里挑 / 剪几条强短视频, start with
standard-clipping.
- If the user says 给同事交接 / 多版本 / 完整工程 / 要做流程沉淀, start with
heavy-collab.
Only ask a routing question when two profiles would change the work materially. Otherwise choose the lighter safe profile and record it in project.yaml.
Story Spec Gates
Read references/story-spec-and-gates.md before creating or changing a cut plan.
Minimum gates before rendering or final subtitles:
- purpose: what the video should make the viewer remember or do
- audience: concrete viewer and viewing context
- platform and target duration
- cold-viewer promise and core information
- source/transcript/material state
- must-keep and must-remove constraints
- visual/subtitle/cover expectations
- references and anti-examples when style matters
If a gate is missing and cannot be inferred from existing project files, ask only the missing high-impact question. Do not make up story goals or visual references silently.
Iteration Discipline
Classify requested changes by impact:
- Light change: wording, subtitle cleanup, accent color, cover title text, small timing adjustment within about 0.5s. Confirm the exact replacement and update the relevant file.
- Medium change: add/drop/reorder a segment, replace PIP, move subtitles, change cover frame, change transition or audio treatment. Update
cut-plan.md, edit-notes.md, and the next version log entry.
- Heavy change: core topic, audience, target platform, target duration, story arc, or the emotional peak changes. Produce an affected-scope summary first, then revise the cut plan after confirmation.
When timecoded feedback exists, preserve the reviewer wording and translate it into the smallest concrete file/version change that fixes the issue.
Standard Outputs
For the exact project folder contract and delivery package, read references/project-structure-and-outputs.md.
final-vNN-*.mp4
cover-vNN-*.jpg|png
subtitles-vNN-*.srt
edit-notes-vNN-*.md
qa-vNN-*.md
project.yaml
cut-plan.md
Workflow
1. Ingest
Create a project folder, fill templates/project.yaml, and follow references/project-structure-and-outputs.md.
Run ffprobe to record source dimensions, duration, frame rate, and audio presence.
2. Transcribe
For Chinese video, use StepFun / StepAudio first unless the user explicitly asks for another provider.
Before ASR, check credentials in this order:
- macOS Keychain service
STEPFUN_API_KEY
- environment variable
STEPFUN_API_KEY
- environment variable
STEP_API_KEY
Never ask the user to paste the raw key into chat, project files, scripts, logs, or GitHub. If StepFun is unavailable, stop and ask before using local ASR such as Whisper.
Before ASR, create a glossary:
- people, gender pronouns, aliases
- book/product/course names
- locations and organization names
- likely ASR confusions
Outputs:
notes/raw-transcript.json
notes/raw-transcript.srt
notes/glossary.md
notes/asr-issues.md
3. Find Story Windows
Search by keywords and read semantically. Do not trust prop visibility alone.
Create notes/candidate-windows.md with:
- source start/end
- topic
- why it matters
- missing context
- emotional/commercial value
- keep/drop recommendation
For requests like “挑 4 条很强的短视频”, start with a candidate pool and selection criteria. Do not pretend the exact clips are known before transcript coverage is sufficient.
4. Build The Cut Plan
Use templates/cut-plan.md and the gates in references/story-spec-and-gates.md.
For 5-10 minute edits, default structure:
- Hook: strongest claim, conflict, or vivid scene
- Context repair: who/what this is and why it matters
- Stakes: difficulty, contradiction, or real-life pressure
- Turn: moment that changes interpretation
- Evidence: quote, story, product, original text, or scene
- Emotional peak: let emotion breathe
- Restrained close: useful reason to act, not hard sell
Never leave a metaphor without setup, a quote without enough context, or a tearful moment without runway.
5. Clean Subtitles
Read references/subtitle-rules.md. Use scripts/subtitle_qa.py before final render.
Default Chinese subtitle house style:
- correct names, titles, places, pronouns, and domain terms
- remove filler words and stutters that do not carry meaning
- use half-width spaces instead of ordinary Chinese punctuation
- no sentence-final punctuation in final subtitles
- keep visible lines around 16 Chinese characters when possible
6. Package Visuals
Read references/cover-and-overlays.md.
Use HyperFrames when the video needs HTML/CSS captions, picture-in-picture assets, quote cards, or timed motion graphics. Prefer one final encode.
Visual priorities:
- speaker face and emotion first
- subtitle readability second
- PIP and quote cards third
- sales or CTA last
7. Voiceover And Audio
For TTS or voiceover, use StepFun / StepAudio first. Default model: stepaudio-2.5-tts. For warm Chinese cultural narration or short-video voiceover, prefer voice ruyananshi with speed around 1.12, then adjust timing with ffmpeg only if needed.
Do not silently fall back to macOS say, local TTS, or mini models. If StepFun cannot be used, stop and ask for approval before fallback.
8. Version And QA
Read references/versioning-and-handoff.md. Use templates/qa.md.
Minimum QA:
ffprobe: 1080x1920, 30fps unless source requires otherwise, audio present- HyperFrames
lint, validate, and inspect when applicable
- subtitle QA script and manual scan
- screenshots at opening, context repair, PIP moments, emotional peak, and close
- audio spot check if inserted assets include original sound
9. Handoff
Update:
edit-notes.md
versions/version-log.md
qa-vNN-*.md
media/assets/asset-log.md
Reviewer feedback should be timecoded. Use templates/review.md.
References
references/story-spec-and-gates.md: director-style gates, question discipline, and iteration impact rules.references/subtitle-rules.md: subtitle cleanup and QA rules.references/story-editing.md: story structure and segment decision rules.references/project-structure-and-outputs.md: exact folder layout, file naming, and delivery package.references/cover-and-overlays.md: cover, PIP, quote card, and HyperFrames guidance.references/versioning-and-handoff.md: file layout, versioning, and teammate review.references/liukaixin-case-study.md: real example distilled from the Liu Kaixin livestream edit.templates/: copyable project, cut plan, review, QA, version log, and asset log templates.
