Agent Skills
› siddsachar/row-bot
› video_editing_guide
video_editing_guide
GitHub提供视频制作与编辑指南，涵盖演示脚本撰写、浏览器录制、多步骤剪辑（变速/缩放/字幕）、品牌风格定制及质量审查，确保生成符合规范的 MP4 视频。
Trigger Scenarios
用户请求制作视频或演示文稿
用户需要对现有视频进行剪辑、调整时长或添加特效
用户询问视频生成的最佳实践或检查标准
Install
npx skills add siddsachar/row-bot --skill video_editing_guide -g -y
SKILL.md
Frontmatter
{
"icon": "🎬",
"name": "video_editing_guide",
"tags": [],
"tools": [
"video_editing"
],
"description": "Guidance for making videos with video and editing them with edit_video.",
"display_name": "Video Editing Guide"
}
- A video of scenes (a demo, how-to, teaser or explainer: recordings, clips, title cards, a voice, music) is the video tool. A scene without a source is a title card, and it voices each line and lays the music itself: never draw cards, mix sound or run ffmpeg with scripts.
- A demo of a site or app in Row-Bot's browser: write the story first, 3-8 steps with one spoken line each. Start browser_record, then before each step mark a beat with its note (a short title) and its line, do the step, and stay on it about as long as the line takes (the beat says how long); beat=work before a wait, cut for a stretch to leave out. Stop, then video action create with recording=rec-…: one scene per beat with its title, line and voice. The camera follows what was done (leave zoom empty).
- The video is the app, not slides: at most an opening and a closing card (3 s, a few words); titles go over the footage as a scene's card.
- Phone demos: browser_viewport phone before recording (taps show as touches); a 16:9 video sets the phone in a device, 9:16 fills the height.
- Its look and voice are the person's: the brand kit (colour, captions, framing, camera, voice, speed, music mood) sets every video, and a video may choose its own voice and voice_speed. Change them when asked.
- After make, read the checks and look at the review sheet with analyze_image. Fix what's off (a dull or long stretch: trim or speed it up; a line that doesn't match its picture: rewrite it) and make it again; it's done only when both are right.
- edit_video edits videos the person already has: recordings (rec-…), earlier edits (edit-…), files in this chat by name, or workspace paths. It never generates video (that is generate_video, when it is on).
- Put the whole edit in one call: inputs in order (several are joined), then steps in order. Each step's times are seconds on the video as it is after the steps before it (after a trim, 0 is the trim's start).
- Demo videos: trim the dead time, speed up waiting or typing with {op: speed, factor: 2-8, start, end} (it shows a "2×" label unless label is false), add a badge with {op: text}, zoom in on a detail with {op: zoom, region: [x, y, width, height], start, end}, and caption with {op: captions, text: <SRT>}. Regions are fractions of the picture.
- Never slow a voice to fill time (it sounds harsh): write more words, or cut the video shorter to fit what's said.
- A video given a voice is checked when it's made. "Needs attention" is not done: a quiet gap wants a line or a shorter part, and a picture held 10 s (20 s under the voice) wants that part cut shorter, sped up or zoomed. Fix it and make it again.
- The result is an MP4 in this chat with a player; do not paste it again. A long edit carries on in the background with its own progress and Stop: say so, and call edit_video with action status and its edit_id to wait.
- The first edit may ask to download ffmpeg; if the person says no or the chat is on Block, say what happened and stop.
- action frames saves stills (count, or times) as one contact sheet; to look at many pictures or a video's frames, use analyze_image with files and contact_sheet. action save copies a finished edit into the workspace (replacing a file there asks first).
Version History
- aea3219 Current 2026-11-02 23:14
