Agent Skills › Forward-Future/scarlet › matt-talking-head

matt-talking-head

GitHub

用于编辑Matt Berman口播视频,通过“click”计数校准唇音同步,处理镜头推近与屏幕演示画面,并依据导演备注筛选最佳表演片段。

skills/matt-talking-head/SKILL.md Forward-Future/scarlet

Trigger Scenarios

需要检查或修复口播视频中唇音不同步的问题 根据导演语音备注进行视频剪辑和取舍得决定

Install

npx skills add Forward-Future/scarlet --skill matt-talking-head -g -y
More Options

Use without installing

npx skills use Forward-Future/scarlet@matt-talking-head

指定 Agent (Claude Code)

npx skills add Forward-Future/scarlet --skill matt-talking-head -a claude-code -g -y

安装 repo 全部 skill

npx skills add Forward-Future/scarlet --all -g -y

预览 repo 内 skill

npx skills add Forward-Future/scarlet --list

SKILL.md

Frontmatter
{
    "name": "matt-talking-head",
    "category": "Story and pacing",
    "description": "Edit Matt Berman’s recorded talking-head videos in Scarlet. Lip sync checked against his spoken “click, 1 2 3 4 5” count, then a dialogue pass cut to his pace, acting on his notes to the editor, with camera punch-ins and any synchronized screen demo framed around what he shows."
}

Matt’s talking-head edit

My edit is a /dialogue-pass cut my way, with my screen demos framed and my camera punched in. Each part below calls the skill it builds on and says what I change. A bare /matt-talking-head covers the current edit, including camera punch-ins and any screen demos. If I name a pass, source, selection or cut, work just on that. Keep an existing edit’s composition, and keep its sync once you have checked it as below. For separate camera, screen and mic recordings, play the dialogue mic once and carry cuts across all synchronized tracks, so my mouth, my voice and the action I describe still line up.

When cropping my camera, make sure to show my full head and shoulders.

Lip sync

Before any cut, make sure my voice and my mouth are in sync. Near the beginning of a recording I say “click” and count “one, two, three, four, five” as a sync mark. Find it and line up the dialogue mic with my mouth on camera at each word: the hard “c” of “click” and the start of each number should land on the frame my lips make it. Where the camera has its own scratch audio, match its waveform to the mic at the count as well. If a screen recording shows a click at that moment, use it to line up the screen too.

When the mic is off, move it, or the camera, to fix it, and save that as its own edit with the offset and the count as the reason. Then check a few spoken words later in the recording and near the end; if the sync has drifted, fix that too and say where. If you can’t find the count, or can’t tell whether my voice and mouth line up, flag it with where you looked rather than guess. Check the same way on an existing edit before working on it.

The count is for sync only and never plays in the finished video: cut it with the wait around it, like a note.

Choose the take

Keep the delivery meant for the viewer. If the wording of two takes is the same but the splice changes which performance survives, keep the take I meant. Keep my genuine reactions.

Keep adjacent identical words when both belong in the sentence, as in “She checked in in Boston.” If the delivery makes it unclear whether one is a stumble, bring it to me for review with the full phrase and the word you would remove.

Notes to my editor

I leave notes for my editor, Alex, while I record. I often say them low, well under my speaking voice, right after the thing they are about: “You can get a screenshot of that. Here.” Others are at full voice: “Use the second take.” “Let me show you the second result” is for the viewer. So is an ask to Alex that’s part of the show: said at full voice in the middle of what I’m telling the viewer, it tells them what they’re about to see, as in “Alex, put the other ones on the screen, please, so we can actually compare them.” Keep that line and do what it asks. How to do it is a note for Alex alone: in “Let me throw them side by side to you, Alex. Have your AI put side by side comparisons of the game here,” the first sentence plays and the rest is a note. Look for these notes through the whole recording, and treat a line said much quieter than the speech around it as a likely note.

Do what a note asks whenever Scarlet can: show the screenshot or the part of the screen I point at, put the two things I name side by side, use the take I ask for, cut what I say to cut. Say what you did in that edit’s summary and quote my words. A note you can’t carry out, such as an opening montage or a music cue, goes in your reply to me with my words and where I said them.

A note never plays in the finished video. Cut a clear one with its lead-in and the wait around it, and don’t flag it. If you can’t tell whether a line is a note or meant for the viewer, keep it and flag it with the doubt. Where a note runs into narration, cut the note alone and flag the join.

My pace

I cut tight, and how tight depends on where the cut falls:

  • A cut inside a sentence leaves next to no space. The next word follows my last as if I had said them in one breath, about five hundredths of a second from word to word.
  • A cut at the end of a sentence leaves a very small space before the next one, about a tenth of a second from my last word to my next.
  • Where no cut falls, leave my own pauses as I made them. A moment while I think, or a beat I hold for emphasis, is how I talk: up to about a second stays. A pause of about a second or more with nothing in it is dead air: it comes out, and its cut follows the two rules above.

Every cut is a jump on my camera, so cut where an editor would: at the mistakes and the dead air. Between them, let me play as I said it. The video should sound like me talking, never chopped up. Cuts that land one after another play as a stutter: a pause beside another cut goes out in that cut or stays, never in a cut of its own.

The two spacings above hold at every cut, one you place by its seconds included, on both sides of it: my last word ends and the cut follows, and after the cut my voice comes straight in. Once I finish a sentence, cut before I look away from the camera: that glance never plays. A change of subject, a new demo or a reaction gets no quiet of its own at a cut on my camera: the jump already marks it, and quiet on my face before or after the cut plays as empty space. A reveal or a result gets the time it earns on the screen, where there is something to look at.

Cut silent typing, searching and loading to the result, and hold the result long enough to see what changed. Flag any silence of two seconds or more that you keep in full, with what the viewer sees or hears during it.

Use a two-frame audio crossfade at dialogue joins and a straight picture cut. Bring my voice in the moment the picture cuts: about two frames from the cut to my voice, as my published cuts do, whether the line starts a sentence, carries one on or opens a new topic, and right on the first sound of a word that starts softly. Where the screen plays across the cut and shows a result that needs a longer look, hold it before my voice; on my camera, keep only a breath that is part of how I say the line. Keep my timing where my voice already enters naturally. Protect the first consonant and the release of the outgoing word through the crossfade.

Show the demo

/frame-screen, with these changes.

During a synchronized screen demo, my camera is an inset in the bottom right, at one size and in that corner throughout. Frame the screen around it: what I talk about, read or use sits clear of the bottom right, and ideally other text does too. The camera doesn't move to make room; the screen does. Don't pick another corner for a demo from the look of the recording: only once you have framed a stretch and seen, with my camera in the bottom right, that no framing keeps what I'm showing clear of it, move the camera for that stretch alone, and say in that edit's reason what it would have covered. Start with the inset size you would otherwise use and scale it up by 20% in both dimensions, then crop the sides for a narrower frame. Aim for 1.5:1 (width to height), with room to go up to 2:1 if the framing needs it. Within the inset, frame me from clear room above my hair down to the desk in front of me, with my full chest and my hands in and some of the room on either side of me, the same throughout the demo. The room above my hair holds when I sit up or lean back: my hair never touches the inset's top edge. When the camera itself leaves little room above my head, keep the whole top of its picture and crop only the bottom and sides. Round all four corners of the inset.

Show the whole of what I am presenting with the rest of the page it is on: the whole watch on its web page, not a close-up of its movement. When framing my screen, keep the browser's own interface (tabs, address bar, bookmarks bar) out of frame wherever the framing allows, and show only the page content. When the page is taller than the video, crop into it; cropping is fine. A move every ten seconds or so is fine while I go from one thing to the next.

Camera punch-ins

/punch-in at every jump cut between sentences on full-screen camera footage. Keep the camera framing steady through mid-sentence cuts. Punch in or out mid-sentence only when my body position or placement in the frame changes enough to make the framing change necessary. That should be rare. Apply this rule once the dialogue cuts are in and again on any cut I approve later.

When requested

  • Captions. /add-captions in the project style.
  • Slide B-roll. Time each supplied slide or requested reveal to the words that introduce it. For new slide art, use one visual idea, a real screenshot, portrait, logo or headline where it helps, large type and few words. Keep the upper right open in the artwork. On a reveal, hold the setup in place and add only the new element. Put spoken detail and sourcing in the slide notes and check the rendered slide at video size.

Version History

  • b4fc0e1 Current 2026-11-22 10:51

    新增唇音同步(Lip sync)章节,引入基于“click”计数的同步标记方法;更新版本变更摘要说明。

  • dbfdef6 2026-11-22 04:41

    优化了摄像机机位与屏幕演示的构图逻辑,确保摄像机保持固定角落并将屏幕围绕其编排;修正了剪辑后的音频衔接问题,消除不必要的静音片段,使声音紧随剪辑点进入。

  • 27a27c9 2026-11-21 16:18

    改进了框架保存时的元数据报告,明确记录镜头覆盖区域的内容;优化了摄像机特写的顶部留白规则,防止主持人头部被裁剪。

  • 16505f9 2026-11-21 10:08

    优化了剪辑后的节拍处理逻辑:确保在话题转换、演示或反应发生时,保留适当的静音节拍;修正了声音过渡参数,避免静音被错误截断,提升剪辑流畅度。

  • bb0c550 2026-11-21 03:58

    更新取镜头选择逻辑,将同词不同表现的版本视为有效选项;明确保留面向观众的完整声音编辑请求,优化对编辑指令的识别与执行。

  • 478ded4 2026-11-20 09:29

    修正了关于自然停顿的剪辑规则,恢复保留约1秒以内的思考或强调性停顿,不再将其作为死寂剪除。

  • f4e426c 2026-11-20 03:21

    优化了停顿修剪逻辑,明确移除0.6秒以上的空白停顿,保留更短的停顿和强调节拍,提升了剪辑效率和质量评分。

  • 6e93a9e 2026-11-19 15:04

    优化停顿处理规则,保留约一秒内的短停顿,防止剪辑点过度密集,并与对话剪辑技能保持一致。

  • e8ad579 2026-11-19 08:58

    新增相机内嵌框架构建逻辑,确保显示完整头部和肩部,并继承之前的镜头插入设定。

  • b6bdb58 2026-11-18 20:45

Same Skill Collection

skills/add-captions/SKILL.md
skills/b-roll/SKILL.md
skills/dialogue-pass/SKILL.md
skills/fit-length/SKILL.md
skills/frame-screen/SKILL.md
skills/make-shorts/SKILL.md
skills/motion-graphics/SKILL.md
skills/punch-in/SKILL.md
skills/speed-through-waits/SKILL.md

Metadata

Files
0
Version
b4fc0e1
Hash
05bd3908
Indexed
2026-11-18 20:45

ホーム - Wiki
Copyright © 2011-2026 iteam. Current version is 2.156.0. UTC+08:00, 2026-11-23 00:52
浙ICP备14020137号-1