Communitygithub.com

ai-video-editing

剪口播、教程、访谈、产品演示或测评,拆解已有参考片,或从长访谈选短片。可按实际录音校对专名和重录、按句选 B-roll 和镜头模板、映射剪后字幕、混音并按争议时码审片。可派生垂直栏目,阿桂校园片搭配 agui-horizontal-video。

ai-video-editing 是什么?

ai-video-editing is a Claude Code agent skill that 剪口播、教程、访谈、产品演示或测评,拆解已有参考片,或从长访谈选短片。可按实际录音校对专名和重录、按句选 B-roll 和镜头模板、映射剪后字幕、混音并按争议时码审片。可派生垂直栏目,阿桂校园片搭配 agui-horizontal-video。.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/suhao1314/ai-auto-video-editing/tree/HEAD/skills/ai-video-editing

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

Universal AI Video Editing

First read the current request, the confirmed version, and the most recent rejection reason. Determine what the user wants to achieve now, then decide which step to start from.

Starting a New Project: First Choose an Existing Editing Method

When the user says "use the general template," read the general shot template and talking-head rough cut. Select existing shots based on the purpose of this sentence. For the first appearance of an interviewee, use a lower-third; to compare two states of the same object, use a split-screen; for teaching an operation, use real UI or screen recording. General projects follow the current person, product, and brand. When the original talking-head footage needs trimming, first run a rough cut that only outputs candidates, then package based on the remaining audio. When the user provides a finalized voiceover or specifies locking the original audio, start from sentence-by-sentence storyboarding.

Given a reference video, find time segments similar to the purpose of the current project. Write down the measured timestamps, the actual visible changes in the frame, and the viewer problem I believe this edit solves. Ensure conclusions can be traced back to a specific frame or sentence. When the user asks "how others edit it and where we differ," read the seven editing workflow comparisons and same-shot verification.

For long talking-head recordings that require cut-by-pause editing, subtitle time offsets, or additional WebVTT delivery; when a specific area of the finalized frame must be preserved; or when the client specifies delivery specifications, use the workbench video tools and tested usage to invoke the corresponding tool. Projects with locked original audio start directly from storyboarding. Pure research collection, batch courseware, or web-side preprocessing alternatives are also on the same page.

What the User SaysWhat to Deliver This Time
"Send me the script first" / "Use my original words as much as possible"Deliver only a readable script. Preserve the author's tone, firsthand experiences, and key figures, segmented by auditory structure. Point out facts that need verification and could change the meaning. Wait until the user requests a storyboard or editing before delivering the packaging table.
"Edit this script into an episode"Inventory assets, deliver a sentence-by-sentence timeline, real asset gaps, and items to be downloaded. Let the user clearly see what to shoot and what to source online, enabling a single response to approve the plan, specify download items, and proceed to editing once ready.
"Cut this interview into several clips"Identify independently valid points, stories, or actions within the full transcription. Provide source in/out points, the first sentence, the closing sentence, and necessary context. After the trial listen, check if "he/she/this person/thing" is understandable without the preceding context, and whether the ending fulfills the beginning. Once approved, create a storyboard for each clip.
"Based on the last version, verify the assets and cut directly"Find the corresponding script, audio, storyboard, and approved specific download items; confirm readiness and cut directly. If crucial evidence is missing, only mention the affected sentences and feasible alternatives.
"This section is too empty" / "Replace this with the real new UI"Find the feedback timestamp, talking-head script, and current editable project. First determine whether the issue is selection, information repetition, or layout. Check for other shots in the same project with the same issue. Only fix the affected areas, delivering the changed timestamps and results with the new version.
"Only add sound effects" / "Change the music" / "Make two versions for comparison"Fix the video and talking-head start point, starting from a confirmed sound-effects-only version without music. When only adding sound effects, preserve the existing music arrangement. When changing music, re-mix on an independent music track. If multitracks are missing and the original track already has mixed music, first check the project backup, explaining the recoverable range. For two versions, export the full video with the same talking-head level, clearly indicating the listening differences.
"This version is final; archive it officially"Find the exact file selected by the user, re-read the completed video, verify the hash before and after copying, then register the formal path and re-editable backup. Clean up based on actual permissions and dependencies.

From Receiving the Script to Starting the Edit

  1. First read the assets. Read the project archiving rules and the retrospectives of similar projects. Inventory the given files and existing asset library; check real content, aspect ratio, audio tracks, duration, readability, and usage scope. Search first, then request. If a finished voiceover exists, extract the actual sound, identify any mixed-in music, sound effects, and pauses; mark which track serves as the time reference. When the user requires preserving the original audio, keep the original mix throughout the edit.
  2. Write a sentence-by-sentence plan for the new video. Use the plan and packing list to deliver a sentence-by-sentence comparison. For each sentence, write: video start and end, original words, specific object or change the audience should see, corresponding asset and its source timestamp, composition and action, how to cut out, subtitle highlights, music and sound effect placements. Continuous demonstrations can cover multiple sentences. If no audio exists, use estimated timings; when audio is available, reorder based on actual speech. If the recorded talking-head differs from the script, use the original sound for subtitles; list factual errors separately for the author to decide. For each gap, provide specific alternatives or minimal re-shoot requirements.
  3. Source assets according to the plan. Divide assets into: those obtained online, those created by AI/design, and real assets that must be provided by the user. Reuse existing locally suitable assets first, preserving their original category and source. You may search and view work pages and preview in-site, and note links before confirmation. Before downloading any video, image, BGM, SFX, icon, or proxy file from a website, first list specific items for the user to explicitly approve this batch or specific items. "Start editing," passing the script, agreeing to a platform, or only having search keywords cannot replace download confirmation. Plans and individual downloads can be confirmed in one go. Already approved items can be safely retried; for additions, changes, different work, paid services, or changes in usage rights, re-confirm the changed items. Execute image generation based on the confirmed generation plan; handle paid services, uploads of private references, or extra expenses according to actual permissions.
  4. Wait for all assets to arrive before starting the edit. Review the selected segments, verify sources, available durations, action directions, cropping, and privacy. On the review table, write "ready / needs replacement / needs to be provided by the user." If gaps remain, go back to the user to choose specific alternatives; if the original footage is sufficient, it can be reused across sentences. The confirmed plan and editing authorization can be combined; once the assets listed are ready, execute directly without asking the same question again.
  5. Determine cut points based on sound. Use the motion and sound manual to choose effective actions; place keyframes, sound effects, and subtitles at the actual moment they occur. Schematic diagrams and dynamic data must be distinguished. Respect existing project entry points. Available tools include existing editing projects, FFmpeg, Remotion, or other locally verified backends. For review, output a new version while preserving the re-editable timeline, subtitles, sources, and any required attribution information.
  6. Review the edit and retain lessons learned. Check the actual exported video; provide the user with a playable version and key timestamps. When feedback, delivery, or published version is returned, use retrospective and vertical derivation to write the results of this episode into the project; read it before starting the next edit. For rounds that only deliver a script or plan, record decisions within the project plan; do not fabricate completed videos or audience effects.

Use the reproducible plan example to record a machine-readable timeline. Existing Excel, Markdown, or editing projects can retain human-readable formats, cross-checked against the same fields. When automated checks are needed, run python scripts/check_plan.py project_plan.json. The script only reads the plan and can verify coverage, timing, items to be downloaded, and conditions for starting the edit; it cannot confirm, download, or render on behalf of the user.

For a frozen version of the GitHub method, license, and verifiable original text, see upstream sources. They provide implementation references; network access, download, and publishing permissions are governed by the current project authorization.

When reviewing, start by looking back at the specific shots mentioned, then verify the final exported file. Use the completed video inspection method to check decoding, audio tracks, subtitle timing; when only changing audio, you can also compare the base video stream and each packet's timestamp. Determine whether the selection fits the original sentence, whether actions are understandable, and whether music overpowers the words by watching and listening respectively. When delivering, write down the actual timestamps checked and results from this batch.

相关技能