video-editing
Video Mode editing guidance for agentic storyboard, timeline, caption, audio, render, and publish work through video_* MCP tools.
- License 许可
- Apache-2.0
- 所属仓库
- bravew/Neumar
- 最近更新
- 2026-11-14
- Stars
- ★ 12
- Forks
- 4
01 前言 Front Matter
- name
- video-editing
- description
- Video Mode editing guidance for agentic storyboard, timeline, caption, audio, render, and publish work through video_* MCP tools.
02 资源文件 5
- 参考文档rubric.yaml711 B
- 参考文档tasks/generation-failure.md332 B
- 参考文档tasks/long-montage.md368 B
- 参考文档tasks/plain-trim-bypass.md237 B
- 参考文档tasks/vivid-photo-outro.md369 B
03 SKILL.md 正文
Video Editing
Use this skill whenever you are editing a Video Mode project through the
mcp__video-edit__*, mcp__media__*, or mcp__ffmpeg-processing__* tools.
Operating Rules
- Read current project state before editing. Start with
video_get_project_summary; usevideo_get_scene,video_get_timeline, andvideo_list_assetsfor narrower context. - Use Video Mode tools for every project mutation. Do not use shell or file
editing tools to change
project.json, assets, timeline JSON, or renders. - Read and reversible write operations can proceed directly. Anything that removes content, changes timing substantially, spends money, renders, or publishes needs explicit approval.
- Prefer one clear operation at a time when the user is reviewing edits.
Use
video_propose_timeline_opsfor multi-op plans before applying them. - Never trim mid-word. When transcript timing exists, snap edit boundaries to word or sentence boundaries and preserve 30-200 ms of context unless the user asks for frame-tight cuts.
- Keep audio continuity deliberate. After visual transitions longer than
500 ms, set the clip audio seam to
cutunless the user asks to preserve source audio under the transition. - Subtitles come after structural edits. Make cuts, reframes, overlays, and narration decisions first, then add or refresh captions.
- For generated images or video, use existing project assets as references when the user asks to keep identity, layout, or product appearance stable.
- Surface cost in plain language after generation or render work. Do not stop because of cost unless the user asks for a budget limit.
- Verify before publish. Use render verification and mention any degraded transitions, missing media, loudness issues, or caption drift.
- Preflight storyboard duration against the project template before approval:
ugc-adallows 15 seconds, themed templates allow 60 seconds, andcustomis unlimited. If the confirmed edit is longer, explain the mismatch and usevideo_set_project_templateor trim it beforevideo_approve_storyboard.video_select_templateis only for HTML/Motion gallery templates and does not change this ceiling. - A vivid photo sequence needs a designed visual treatment, not only generic zoom keyframes. Analyze each photo and adjacent footage, choose one coherent vivid-overlay or gallery-template family, vary content-aware motion, and inspect composited frames. Keep the supplied photos recognizable; generate derived or replacement imagery only when the user asks and approves cost.
- If image/video generation fails, report the provider error. Do not silently substitute a placeholder, unrelated asset, or unmaterialized AI scene.
Project directory contract
Every Video Mode project on disk uses this layout. Honor it when picking
work directories for ad-hoc tool calls (e.g. mcp__ffmpeg-processing__*)
and when describing where files live to the user:
<projectDir>/
project.json # journal + storyboard + timeline + asset registry
assets/ # ALL source / working media
# - generated images (Seedream / Imagen / GPT-Image)
# - generated videos (Veo / Seedance / Runway)
# - generated voiceover (TTS) and music
# - generated b-roll, lipsync renders
# - user uploads and reference images
# - any ffmpeg intermediate the agent produces
# The runtime steers media MCP outputs here automatically.
sources/ # imported original captures before processing
cache/ # render cache (per-scene + per-clip previews)
output/ # ONLY the final rendered video + its caption .srt
# Do not write anything else here.
If you call mcp__ffmpeg-processing__* for an ad-hoc transform (trim,
concat, audio normalize, frame extract), pass work_dir as
<projectDir>/assets so the result lands with the rest of the working
media. Reach <projectDir> via video_get_project_summary if you don't
already have it from earlier in the turn.
Tool Selection
video_get_project_summary: first read of every turn.video_set_project_template: change the duration-governing project template; use it for an approved long-form edit that cannot pass the current ceiling.video_describe_scene: user asks what is in a scene or wants an edit scoped to a scene.video_add_scene,video_remove_scene,video_reorder_scenes: storyboard structure changes.video_set_duration,video_set_transition,video_set_timeline_bookend,video_set_clip_audio_seam: timeline timing and boundary polish.video_propose_timeline_ops: batch edits, auto-cuts, pacing changes, source capture insertion, or recipe-driven operation sets.video_apply_timeline_op: one reviewed timeline operation.video_set_caption,video_add_captions: caption content and style.video_generate_voiceover,video_generate_music: narration or music plans.mcp__media__media_generate_imageandmcp__media__media_generate_video: asset generation. Ask before calls because they can spend money. The output file is auto-registered as a project asset by the runtime — you do not need to call any follow-up tool to attach it to the library. Mention to the user that the new asset is in the project's Assets panel and offer to attach it to a specific scene withvideo_attach_asset(with sceneId) when the context implies a target scene.video_analyze_image,video_list_overlay_presets,video_inspect_timeline_frames: the minimum loop for vivid, cinematic, or style-matched photo slides. Prefer a consistent overlay/frame family plus subject-aware motion over three repeated scale-only moves.video_reframe: aspect-ratio adaptation.video_render,video_verify_render,video_publish_to: finalization path.
Approval Copy
When asking for approval, be concrete:
- Say what will change.
- Name the affected scene, clip, track, aspect ratio, or destination.
- Mention likely cost for generation/render work when available.
- Keep the approval text short enough to fit a card.
Good examples:
- "Generate 4 reference images for Scene 2 using the product hero asset?"
- "Trim clip
clip-voice-1from 12.4s to 18.9s and journal undo data?" - "Render a 9:16 draft locally with burned-in captions?"
Response Discipline
After tool use, summarize the result, not the implementation. If an edit was journaled, mention that undo is available. If a requested operation cannot be done with the current tools, say which project fact is missing and ask for the smallest next input.
04 同仓库更多 Skills
yt-dlp
A robust CLI wrapper for yt-dlp to download videos, playlists, and audio from YouTube and thousands of other sites. Supports format selection, quality control, metadata embedding, and cookie authentication.
weather
Get current weather and forecasts (no API key required).
video-edit
Conversational video editing workflow for cuts, transcript-based edits, overlays, captions, subtitle burn-in, and render planning with ffmpeg. Use when the user asks to edit video by instruction, cut silence or words, add subtitles, add overlays, transcribe video, or render a final clip.
skill-creator
Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, update or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
ship
Full shipping workflow: test, review, version, changelog, and PR creation. Use when: preparing to ship code, creating releases, merging PRs.
security-audit
Application security audit covering OWASP Top 10, STRIDE threat modeling, and secrets archaeology. Use when: security review, pre-launch audit, dependency scanning.